2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2004 Poul-Henning Kamp
5 * Copyright (c) 1994,1997 John S. Dyson
6 * Copyright (c) 2013 The FreeBSD Foundation
9 * Portions of this software were developed by Konstantin Belousov
10 * under sponsorship from the FreeBSD Foundation.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * this file contains a new buffer I/O scheme implementing a coherent
36 * VM object and buffer cache scheme. Pains have been taken to make
37 * sure that the performance degradation associated with schemes such
38 * as this is not realized.
40 * Author: John S. Dyson
41 * Significant help during the development and debugging phases
42 * had been provided by David Greenman, also of the FreeBSD core team.
44 * see man buf(9) for more info.
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
50 #include <sys/param.h>
51 #include <sys/systm.h>
53 #include <sys/bitset.h>
55 #include <sys/counter.h>
57 #include <sys/devicestat.h>
58 #include <sys/eventhandler.h>
61 #include <sys/limits.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
69 #include <sys/racct.h>
70 #include <sys/refcount.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
74 #include <sys/sysctl.h>
75 #include <sys/sysproto.h>
77 #include <sys/vmmeter.h>
78 #include <sys/vnode.h>
79 #include <sys/watchdog.h>
80 #include <geom/geom.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/swap_pager.h>
92 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
94 struct bio_ops bioops; /* I/O operation notification */
96 struct buf_ops buf_ops_bio = {
97 .bop_name = "buf_ops_bio",
98 .bop_write = bufwrite,
99 .bop_strategy = bufstrategy,
101 .bop_bdflush = bufbdflush,
105 struct mtx_padalign bq_lock;
106 TAILQ_HEAD(, buf) bq_queue;
108 uint16_t bq_subqueue;
110 } __aligned(CACHE_LINE_SIZE);
112 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
113 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
114 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
115 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
118 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
119 struct bufqueue bd_dirtyq;
120 struct bufqueue *bd_cleanq;
121 struct mtx_padalign bd_run_lock;
126 long bd_bufspacethresh;
127 int bd_hifreebuffers;
128 int bd_lofreebuffers;
129 int bd_hidirtybuffers;
130 int bd_lodirtybuffers;
131 int bd_dirtybufthresh;
135 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
136 int __aligned(CACHE_LINE_SIZE) bd_running;
137 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
138 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
139 } __aligned(CACHE_LINE_SIZE);
141 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
142 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
143 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
144 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
145 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
146 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
147 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
148 #define BD_DOMAIN(bd) (bd - bdomain)
150 static struct buf *buf; /* buffer header pool */
151 extern struct buf *swbuf; /* Swap buffer header pool. */
152 caddr_t unmapped_buf;
154 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
155 struct proc *bufdaemonproc;
157 static int inmem(struct vnode *vp, daddr_t blkno);
158 static void vm_hold_free_pages(struct buf *bp, int newbsize);
159 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
161 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
162 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
164 static void vfs_clean_pages_dirty_buf(struct buf *bp);
165 static void vfs_setdirty_range(struct buf *bp);
166 static void vfs_vmio_invalidate(struct buf *bp);
167 static void vfs_vmio_truncate(struct buf *bp, int npages);
168 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
169 static int vfs_bio_clcheck(struct vnode *vp, int size,
170 daddr_t lblkno, daddr_t blkno);
171 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
172 void (*)(struct buf *));
173 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
174 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
175 static void buf_daemon(void);
176 static __inline void bd_wakeup(void);
177 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
178 static void bufkva_reclaim(vmem_t *, int);
179 static void bufkva_free(struct buf *);
180 static int buf_import(void *, void **, int, int, int);
181 static void buf_release(void *, void **, int);
182 static void maxbcachebuf_adjust(void);
183 static inline struct bufdomain *bufdomain(struct buf *);
184 static void bq_remove(struct bufqueue *bq, struct buf *bp);
185 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
186 static int buf_recycle(struct bufdomain *, bool kva);
187 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
188 const char *lockname);
189 static void bd_init(struct bufdomain *bd);
190 static int bd_flushall(struct bufdomain *bd);
191 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
192 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
194 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
195 int vmiodirenable = TRUE;
196 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
197 "Use the VM system for directory writes");
198 long runningbufspace;
199 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
200 "Amount of presently outstanding async buffer io");
201 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
202 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
203 static counter_u64_t bufkvaspace;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
205 "Kernel virtual memory used for buffers");
206 static long maxbufspace;
207 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
208 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
209 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
210 "Maximum allowed value of bufspace (including metadata)");
211 static long bufmallocspace;
212 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
213 "Amount of malloced memory for buffers");
214 static long maxbufmallocspace;
215 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
216 0, "Maximum amount of malloced memory for buffers");
217 static long lobufspace;
218 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
219 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
220 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
221 "Minimum amount of buffers we want to have");
223 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
224 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
225 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
226 "Maximum allowed value of bufspace (excluding metadata)");
228 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
229 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
230 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
231 "Bufspace consumed before waking the daemon to free some");
232 static counter_u64_t buffreekvacnt;
233 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
234 "Number of times we have freed the KVA space from some buffer");
235 static counter_u64_t bufdefragcnt;
236 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
237 "Number of times we have had to repeat buffer allocation to defragment");
238 static long lorunningspace;
239 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
240 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
241 "Minimum preferred space used for in-progress I/O");
242 static long hirunningspace;
243 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
244 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
245 "Maximum amount of space to use for in-progress I/O");
246 int dirtybufferflushes;
247 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
248 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
250 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
251 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
252 int altbufferflushes;
253 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
254 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
255 static int recursiveflushes;
256 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
257 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
258 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
259 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
260 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
261 "Number of buffers that are dirty (has unwritten changes) at the moment");
262 static int lodirtybuffers;
263 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
264 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
265 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
266 "How many buffers we want to have free before bufdaemon can sleep");
267 static int hidirtybuffers;
268 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
269 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
270 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
271 "When the number of dirty buffers is considered severe");
273 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
274 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
275 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
276 "Number of bdwrite to bawrite conversions to clear dirty buffers");
277 static int numfreebuffers;
278 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
279 "Number of free buffers");
280 static int lofreebuffers;
281 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
282 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
283 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
284 "Target number of free buffers");
285 static int hifreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
287 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
288 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
289 "Threshold for clean buffer recycling");
290 static counter_u64_t getnewbufcalls;
291 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
292 &getnewbufcalls, "Number of calls to getnewbuf");
293 static counter_u64_t getnewbufrestarts;
294 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
296 "Number of times getnewbuf has had to restart a buffer acquisition");
297 static counter_u64_t mappingrestarts;
298 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
300 "Number of times getblk has had to restart a buffer mapping for "
302 static counter_u64_t numbufallocfails;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
304 &numbufallocfails, "Number of times buffer allocations failed");
305 static int flushbufqtarget = 100;
306 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
307 "Amount of work to do in flushbufqueues when helping bufdaemon");
308 static counter_u64_t notbufdflushes;
309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
310 "Number of dirty buffer flushes done by the bufdaemon helpers");
311 static long barrierwrites;
312 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
313 &barrierwrites, 0, "Number of barrier writes");
314 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
315 &unmapped_buf_allowed, 0,
316 "Permit the use of the unmapped i/o");
317 int maxbcachebuf = MAXBCACHEBUF;
318 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
319 "Maximum size of a buffer cache block");
322 * This lock synchronizes access to bd_request.
324 static struct mtx_padalign __exclusive_cache_line bdlock;
327 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
328 * waitrunningbufspace().
330 static struct mtx_padalign __exclusive_cache_line rbreqlock;
333 * Lock that protects bdirtywait.
335 static struct mtx_padalign __exclusive_cache_line bdirtylock;
338 * Wakeup point for bufdaemon, as well as indicator of whether it is already
339 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
342 static int bd_request;
345 * Request for the buf daemon to write more buffers than is indicated by
346 * lodirtybuf. This may be necessary to push out excess dependencies or
347 * defragment the address space where a simple count of the number of dirty
348 * buffers is insufficient to characterize the demand for flushing them.
350 static int bd_speedupreq;
353 * Synchronization (sleep/wakeup) variable for active buffer space requests.
354 * Set when wait starts, cleared prior to wakeup().
355 * Used in runningbufwakeup() and waitrunningbufspace().
357 static int runningbufreq;
360 * Synchronization for bwillwrite() waiters.
362 static int bdirtywait;
365 * Definitions for the buffer free lists.
367 #define QUEUE_NONE 0 /* on no queue */
368 #define QUEUE_EMPTY 1 /* empty buffer headers */
369 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
370 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
371 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
373 /* Maximum number of buffer domains. */
374 #define BUF_DOMAINS 8
376 struct bufdomainset bdlodirty; /* Domains > lodirty */
377 struct bufdomainset bdhidirty; /* Domains > hidirty */
379 /* Configured number of clean queues. */
380 static int __read_mostly buf_domains;
382 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
383 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
384 struct bufqueue __exclusive_cache_line bqempty;
387 * per-cpu empty buffer cache.
392 * Single global constant for BUF_WMESG, to avoid getting multiple references.
393 * buf_wmesg is referred from macros.
395 const char *buf_wmesg = BUF_WMESG;
398 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
403 value = *(long *)arg1;
404 error = sysctl_handle_long(oidp, &value, 0, req);
405 if (error != 0 || req->newptr == NULL)
407 mtx_lock(&rbreqlock);
408 if (arg1 == &hirunningspace) {
409 if (value < lorunningspace)
412 hirunningspace = value;
414 KASSERT(arg1 == &lorunningspace,
415 ("%s: unknown arg1", __func__));
416 if (value > hirunningspace)
419 lorunningspace = value;
421 mtx_unlock(&rbreqlock);
426 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
432 value = *(int *)arg1;
433 error = sysctl_handle_int(oidp, &value, 0, req);
434 if (error != 0 || req->newptr == NULL)
436 *(int *)arg1 = value;
437 for (i = 0; i < buf_domains; i++)
438 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
445 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
451 value = *(long *)arg1;
452 error = sysctl_handle_long(oidp, &value, 0, req);
453 if (error != 0 || req->newptr == NULL)
455 *(long *)arg1 = value;
456 for (i = 0; i < buf_domains; i++)
457 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
463 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
464 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
466 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
473 for (i = 0; i < buf_domains; i++)
474 lvalue += bdomain[i].bd_bufspace;
475 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
476 return (sysctl_handle_long(oidp, &lvalue, 0, req));
477 if (lvalue > INT_MAX)
478 /* On overflow, still write out a long to trigger ENOMEM. */
479 return (sysctl_handle_long(oidp, &lvalue, 0, req));
481 return (sysctl_handle_int(oidp, &ivalue, 0, req));
485 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
491 for (i = 0; i < buf_domains; i++)
492 lvalue += bdomain[i].bd_bufspace;
493 return (sysctl_handle_long(oidp, &lvalue, 0, req));
498 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
504 for (i = 0; i < buf_domains; i++)
505 value += bdomain[i].bd_numdirtybuffers;
506 return (sysctl_handle_int(oidp, &value, 0, req));
512 * Wakeup any bwillwrite() waiters.
517 mtx_lock(&bdirtylock);
522 mtx_unlock(&bdirtylock);
528 * Clear a domain from the appropriate bitsets when dirtybuffers
532 bd_clear(struct bufdomain *bd)
535 mtx_lock(&bdirtylock);
536 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
537 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
538 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
539 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
540 mtx_unlock(&bdirtylock);
546 * Set a domain in the appropriate bitsets when dirtybuffers
550 bd_set(struct bufdomain *bd)
553 mtx_lock(&bdirtylock);
554 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
555 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
556 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
557 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
558 mtx_unlock(&bdirtylock);
564 * Decrement the numdirtybuffers count by one and wakeup any
565 * threads blocked in bwillwrite().
568 bdirtysub(struct buf *bp)
570 struct bufdomain *bd;
574 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
575 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
577 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
584 * Increment the numdirtybuffers count by one and wakeup the buf
588 bdirtyadd(struct buf *bp)
590 struct bufdomain *bd;
594 * Only do the wakeup once as we cross the boundary. The
595 * buf daemon will keep running until the condition clears.
598 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
599 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
601 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
606 * bufspace_daemon_wakeup:
608 * Wakeup the daemons responsible for freeing clean bufs.
611 bufspace_daemon_wakeup(struct bufdomain *bd)
615 * avoid the lock if the daemon is running.
617 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
619 atomic_store_int(&bd->bd_running, 1);
620 wakeup(&bd->bd_running);
626 * bufspace_daemon_wait:
628 * Sleep until the domain falls below a limit or one second passes.
631 bufspace_daemon_wait(struct bufdomain *bd)
634 * Re-check our limits and sleep. bd_running must be
635 * cleared prior to checking the limits to avoid missed
636 * wakeups. The waker will adjust one of bufspace or
637 * freebuffers prior to checking bd_running.
640 atomic_store_int(&bd->bd_running, 0);
641 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
642 bd->bd_freebuffers > bd->bd_lofreebuffers) {
643 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
646 /* Avoid spurious wakeups while running. */
647 atomic_store_int(&bd->bd_running, 1);
655 * Adjust the reported bufspace for a KVA managed buffer, possibly
656 * waking any waiters.
659 bufspace_adjust(struct buf *bp, int bufsize)
661 struct bufdomain *bd;
665 KASSERT((bp->b_flags & B_MALLOC) == 0,
666 ("bufspace_adjust: malloc buf %p", bp));
668 diff = bufsize - bp->b_bufsize;
670 atomic_subtract_long(&bd->bd_bufspace, -diff);
671 } else if (diff > 0) {
672 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
673 /* Wake up the daemon on the transition. */
674 if (space < bd->bd_bufspacethresh &&
675 space + diff >= bd->bd_bufspacethresh)
676 bufspace_daemon_wakeup(bd);
678 bp->b_bufsize = bufsize;
684 * Reserve bufspace before calling allocbuf(). metadata has a
685 * different space limit than data.
688 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
694 limit = bd->bd_maxbufspace;
696 limit = bd->bd_hibufspace;
697 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
700 atomic_subtract_long(&bd->bd_bufspace, size);
704 /* Wake up the daemon on the transition. */
705 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
706 bufspace_daemon_wakeup(bd);
714 * Release reserved bufspace after bufspace_adjust() has consumed it.
717 bufspace_release(struct bufdomain *bd, int size)
720 atomic_subtract_long(&bd->bd_bufspace, size);
726 * Wait for bufspace, acting as the buf daemon if a locked vnode is
727 * supplied. bd_wanted must be set prior to polling for space. The
728 * operation must be re-tried on return.
731 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
732 int slpflag, int slptimeo)
735 int error, fl, norunbuf;
737 if ((gbflags & GB_NOWAIT_BD) != 0)
742 while (bd->bd_wanted) {
743 if (vp != NULL && vp->v_type != VCHR &&
744 (td->td_pflags & TDP_BUFNEED) == 0) {
747 * getblk() is called with a vnode locked, and
748 * some majority of the dirty buffers may as
749 * well belong to the vnode. Flushing the
750 * buffers there would make a progress that
751 * cannot be achieved by the buf_daemon, that
752 * cannot lock the vnode.
754 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
755 (td->td_pflags & TDP_NORUNNINGBUF);
758 * Play bufdaemon. The getnewbuf() function
759 * may be called while the thread owns lock
760 * for another dirty buffer for the same
761 * vnode, which makes it impossible to use
762 * VOP_FSYNC() there, due to the buffer lock
765 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
766 fl = buf_flush(vp, bd, flushbufqtarget);
767 td->td_pflags &= norunbuf;
771 if (bd->bd_wanted == 0)
774 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
775 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
786 * buffer space management daemon. Tries to maintain some marginal
787 * amount of free buffer space so that requesting processes neither
788 * block nor work to reclaim buffers.
791 bufspace_daemon(void *arg)
793 struct bufdomain *bd;
795 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
796 SHUTDOWN_PRI_LAST + 100);
800 kthread_suspend_check();
803 * Free buffers from the clean queue until we meet our
806 * Theory of operation: The buffer cache is most efficient
807 * when some free buffer headers and space are always
808 * available to getnewbuf(). This daemon attempts to prevent
809 * the excessive blocking and synchronization associated
810 * with shortfall. It goes through three phases according
813 * 1) The daemon wakes up voluntarily once per-second
814 * during idle periods when the counters are below
815 * the wakeup thresholds (bufspacethresh, lofreebuffers).
817 * 2) The daemon wakes up as we cross the thresholds
818 * ahead of any potential blocking. This may bounce
819 * slightly according to the rate of consumption and
822 * 3) The daemon and consumers are starved for working
823 * clean buffers. This is the 'bufspace' sleep below
824 * which will inefficiently trade bufs with bqrelse
825 * until we return to condition 2.
827 while (bd->bd_bufspace > bd->bd_lobufspace ||
828 bd->bd_freebuffers < bd->bd_hifreebuffers) {
829 if (buf_recycle(bd, false) != 0) {
833 * Speedup dirty if we've run out of clean
834 * buffers. This is possible in particular
835 * because softdep may held many bufs locked
836 * pending writes to other bufs which are
837 * marked for delayed write, exhausting
838 * clean space until they are written.
843 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
844 PRIBIO|PDROP, "bufspace", hz/10);
850 bufspace_daemon_wait(bd);
857 * Adjust the reported bufspace for a malloc managed buffer, possibly
858 * waking any waiters.
861 bufmallocadjust(struct buf *bp, int bufsize)
865 KASSERT((bp->b_flags & B_MALLOC) != 0,
866 ("bufmallocadjust: non-malloc buf %p", bp));
867 diff = bufsize - bp->b_bufsize;
869 atomic_subtract_long(&bufmallocspace, -diff);
871 atomic_add_long(&bufmallocspace, diff);
872 bp->b_bufsize = bufsize;
878 * Wake up processes that are waiting on asynchronous writes to fall
879 * below lorunningspace.
885 mtx_lock(&rbreqlock);
888 wakeup(&runningbufreq);
890 mtx_unlock(&rbreqlock);
896 * Decrement the outstanding write count according.
899 runningbufwakeup(struct buf *bp)
903 bspace = bp->b_runningbufspace;
906 space = atomic_fetchadd_long(&runningbufspace, -bspace);
907 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
909 bp->b_runningbufspace = 0;
911 * Only acquire the lock and wakeup on the transition from exceeding
912 * the threshold to falling below it.
914 if (space < lorunningspace)
916 if (space - bspace > lorunningspace)
922 * waitrunningbufspace()
924 * runningbufspace is a measure of the amount of I/O currently
925 * running. This routine is used in async-write situations to
926 * prevent creating huge backups of pending writes to a device.
927 * Only asynchronous writes are governed by this function.
929 * This does NOT turn an async write into a sync write. It waits
930 * for earlier writes to complete and generally returns before the
931 * caller's write has reached the device.
934 waitrunningbufspace(void)
937 mtx_lock(&rbreqlock);
938 while (runningbufspace > hirunningspace) {
940 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
942 mtx_unlock(&rbreqlock);
947 * vfs_buf_test_cache:
949 * Called when a buffer is extended. This function clears the B_CACHE
950 * bit if the newly extended portion of the buffer does not contain
954 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
955 vm_offset_t size, vm_page_t m)
959 * This function and its results are protected by higher level
960 * synchronization requiring vnode and buf locks to page in and
963 if (bp->b_flags & B_CACHE) {
964 int base = (foff + off) & PAGE_MASK;
965 if (vm_page_is_valid(m, base, size) == 0)
966 bp->b_flags &= ~B_CACHE;
970 /* Wake up the buffer daemon if necessary */
976 if (bd_request == 0) {
984 * Adjust the maxbcachbuf tunable.
987 maxbcachebuf_adjust(void)
992 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
995 while (i * 2 <= maxbcachebuf)
998 if (maxbcachebuf < MAXBSIZE)
999 maxbcachebuf = MAXBSIZE;
1000 if (maxbcachebuf > MAXPHYS)
1001 maxbcachebuf = MAXPHYS;
1002 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1003 printf("maxbcachebuf=%d\n", maxbcachebuf);
1007 * bd_speedup - speedup the buffer cache flushing code
1016 if (bd_speedupreq == 0 || bd_request == 0)
1021 wakeup(&bd_request);
1022 mtx_unlock(&bdlock);
1026 #define TRANSIENT_DENOM 5
1028 #define TRANSIENT_DENOM 10
1032 * Calculating buffer cache scaling values and reserve space for buffer
1033 * headers. This is called during low level kernel initialization and
1034 * may be called more then once. We CANNOT write to the memory area
1035 * being reserved at this time.
1038 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1041 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1044 * physmem_est is in pages. Convert it to kilobytes (assumes
1045 * PAGE_SIZE is >= 1K)
1047 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1049 maxbcachebuf_adjust();
1051 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1052 * For the first 64MB of ram nominally allocate sufficient buffers to
1053 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1054 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1055 * the buffer cache we limit the eventual kva reservation to
1058 * factor represents the 1/4 x ram conversion.
1061 int factor = 4 * BKVASIZE / 1024;
1064 if (physmem_est > 4096)
1065 nbuf += min((physmem_est - 4096) / factor,
1067 if (physmem_est > 65536)
1068 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1069 32 * 1024 * 1024 / (factor * 5));
1071 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1072 nbuf = maxbcache / BKVASIZE;
1077 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1078 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1079 if (nbuf > maxbuf) {
1081 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1087 * Ideal allocation size for the transient bio submap is 10%
1088 * of the maximal space buffer map. This roughly corresponds
1089 * to the amount of the buffer mapped for typical UFS load.
1091 * Clip the buffer map to reserve space for the transient
1092 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1093 * maximum buffer map extent on the platform.
1095 * The fall-back to the maxbuf in case of maxbcache unset,
1096 * allows to not trim the buffer KVA for the architectures
1097 * with ample KVA space.
1099 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1100 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1101 buf_sz = (long)nbuf * BKVASIZE;
1102 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1103 (TRANSIENT_DENOM - 1)) {
1105 * There is more KVA than memory. Do not
1106 * adjust buffer map size, and assign the rest
1107 * of maxbuf to transient map.
1109 biotmap_sz = maxbuf_sz - buf_sz;
1112 * Buffer map spans all KVA we could afford on
1113 * this platform. Give 10% (20% on i386) of
1114 * the buffer map to the transient bio map.
1116 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1117 buf_sz -= biotmap_sz;
1119 if (biotmap_sz / INT_MAX > MAXPHYS)
1120 bio_transient_maxcnt = INT_MAX;
1122 bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1124 * Artificially limit to 1024 simultaneous in-flight I/Os
1125 * using the transient mapping.
1127 if (bio_transient_maxcnt > 1024)
1128 bio_transient_maxcnt = 1024;
1130 nbuf = buf_sz / BKVASIZE;
1134 nswbuf = min(nbuf / 4, 256);
1135 if (nswbuf < NSWBUF_MIN)
1136 nswbuf = NSWBUF_MIN;
1140 * Reserve space for the buffer cache buffers
1143 v = (caddr_t)(buf + nbuf);
1148 /* Initialize the buffer subsystem. Called before use of any buffers. */
1155 KASSERT(maxbcachebuf >= MAXBSIZE,
1156 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1158 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1159 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1160 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1161 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1163 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1165 /* finally, initialize each buffer header and stick on empty q */
1166 for (i = 0; i < nbuf; i++) {
1168 bzero(bp, sizeof *bp);
1169 bp->b_flags = B_INVAL;
1170 bp->b_rcred = NOCRED;
1171 bp->b_wcred = NOCRED;
1172 bp->b_qindex = QUEUE_NONE;
1174 bp->b_subqueue = mp_maxid + 1;
1176 bp->b_data = bp->b_kvabase = unmapped_buf;
1177 LIST_INIT(&bp->b_dep);
1179 bq_insert(&bqempty, bp, false);
1183 * maxbufspace is the absolute maximum amount of buffer space we are
1184 * allowed to reserve in KVM and in real terms. The absolute maximum
1185 * is nominally used by metadata. hibufspace is the nominal maximum
1186 * used by most other requests. The differential is required to
1187 * ensure that metadata deadlocks don't occur.
1189 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1190 * this may result in KVM fragmentation which is not handled optimally
1191 * by the system. XXX This is less true with vmem. We could use
1194 maxbufspace = (long)nbuf * BKVASIZE;
1195 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1196 lobufspace = (hibufspace / 20) * 19; /* 95% */
1197 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1200 * Note: The 16 MiB upper limit for hirunningspace was chosen
1201 * arbitrarily and may need further tuning. It corresponds to
1202 * 128 outstanding write IO requests (if IO size is 128 KiB),
1203 * which fits with many RAID controllers' tagged queuing limits.
1204 * The lower 1 MiB limit is the historical upper limit for
1207 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1208 16 * 1024 * 1024), 1024 * 1024);
1209 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1212 * Limit the amount of malloc memory since it is wired permanently into
1213 * the kernel space. Even though this is accounted for in the buffer
1214 * allocation, we don't want the malloced region to grow uncontrolled.
1215 * The malloc scheme improves memory utilization significantly on
1216 * average (small) directories.
1218 maxbufmallocspace = hibufspace / 20;
1221 * Reduce the chance of a deadlock occurring by limiting the number
1222 * of delayed-write dirty buffers we allow to stack up.
1224 hidirtybuffers = nbuf / 4 + 20;
1225 dirtybufthresh = hidirtybuffers * 9 / 10;
1227 * To support extreme low-memory systems, make sure hidirtybuffers
1228 * cannot eat up all available buffer space. This occurs when our
1229 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1230 * buffer space assuming BKVASIZE'd buffers.
1232 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1233 hidirtybuffers >>= 1;
1235 lodirtybuffers = hidirtybuffers / 2;
1238 * lofreebuffers should be sufficient to avoid stalling waiting on
1239 * buf headers under heavy utilization. The bufs in per-cpu caches
1240 * are counted as free but will be unavailable to threads executing
1243 * hifreebuffers is the free target for the bufspace daemon. This
1244 * should be set appropriately to limit work per-iteration.
1246 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1247 hifreebuffers = (3 * lofreebuffers) / 2;
1248 numfreebuffers = nbuf;
1250 /* Setup the kva and free list allocators. */
1251 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1252 buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1253 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1256 * Size the clean queue according to the amount of buffer space.
1257 * One queue per-256mb up to the max. More queues gives better
1258 * concurrency but less accurate LRU.
1260 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1261 for (i = 0 ; i < buf_domains; i++) {
1262 struct bufdomain *bd;
1266 bd->bd_freebuffers = nbuf / buf_domains;
1267 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1268 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1269 bd->bd_bufspace = 0;
1270 bd->bd_maxbufspace = maxbufspace / buf_domains;
1271 bd->bd_hibufspace = hibufspace / buf_domains;
1272 bd->bd_lobufspace = lobufspace / buf_domains;
1273 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1274 bd->bd_numdirtybuffers = 0;
1275 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1276 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1277 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1278 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1279 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1281 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1282 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1283 mappingrestarts = counter_u64_alloc(M_WAITOK);
1284 numbufallocfails = counter_u64_alloc(M_WAITOK);
1285 notbufdflushes = counter_u64_alloc(M_WAITOK);
1286 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1287 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1288 bufkvaspace = counter_u64_alloc(M_WAITOK);
1293 vfs_buf_check_mapped(struct buf *bp)
1296 KASSERT(bp->b_kvabase != unmapped_buf,
1297 ("mapped buf: b_kvabase was not updated %p", bp));
1298 KASSERT(bp->b_data != unmapped_buf,
1299 ("mapped buf: b_data was not updated %p", bp));
1300 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1301 MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1305 vfs_buf_check_unmapped(struct buf *bp)
1308 KASSERT(bp->b_data == unmapped_buf,
1309 ("unmapped buf: corrupted b_data %p", bp));
1312 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1313 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1315 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1316 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1320 isbufbusy(struct buf *bp)
1322 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1323 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1329 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1332 bufshutdown(int show_busybufs)
1334 static int first_buf_printf = 1;
1336 int iter, nbusy, pbusy;
1342 * Sync filesystems for shutdown
1344 wdog_kern_pat(WD_LASTVAL);
1345 sys_sync(curthread, NULL);
1348 * With soft updates, some buffers that are
1349 * written will be remarked as dirty until other
1350 * buffers are written.
1352 for (iter = pbusy = 0; iter < 20; iter++) {
1354 for (bp = &buf[nbuf]; --bp >= buf; )
1358 if (first_buf_printf)
1359 printf("All buffers synced.");
1362 if (first_buf_printf) {
1363 printf("Syncing disks, buffers remaining... ");
1364 first_buf_printf = 0;
1366 printf("%d ", nbusy);
1371 wdog_kern_pat(WD_LASTVAL);
1372 sys_sync(curthread, NULL);
1376 * Spin for a while to allow interrupt threads to run.
1378 DELAY(50000 * iter);
1381 * Context switch several times to allow interrupt
1384 for (subiter = 0; subiter < 50 * iter; subiter++) {
1385 thread_lock(curthread);
1386 mi_switch(SW_VOL, NULL);
1387 thread_unlock(curthread);
1394 * Count only busy local buffers to prevent forcing
1395 * a fsck if we're just a client of a wedged NFS server
1398 for (bp = &buf[nbuf]; --bp >= buf; ) {
1399 if (isbufbusy(bp)) {
1401 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1402 if (bp->b_dev == NULL) {
1403 TAILQ_REMOVE(&mountlist,
1404 bp->b_vp->v_mount, mnt_list);
1409 if (show_busybufs > 0) {
1411 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1412 nbusy, bp, bp->b_vp, bp->b_flags,
1413 (intmax_t)bp->b_blkno,
1414 (intmax_t)bp->b_lblkno);
1415 BUF_LOCKPRINTINFO(bp);
1416 if (show_busybufs > 1)
1424 * Failed to sync all blocks. Indicate this and don't
1425 * unmount filesystems (thus forcing an fsck on reboot).
1427 printf("Giving up on %d buffers\n", nbusy);
1428 DELAY(5000000); /* 5 seconds */
1430 if (!first_buf_printf)
1431 printf("Final sync complete\n");
1433 * Unmount filesystems
1435 if (panicstr == NULL)
1439 DELAY(100000); /* wait for console output to finish */
1443 bpmap_qenter(struct buf *bp)
1446 BUF_CHECK_MAPPED(bp);
1449 * bp->b_data is relative to bp->b_offset, but
1450 * bp->b_offset may be offset into the first page.
1452 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1453 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1454 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1455 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1458 static inline struct bufdomain *
1459 bufdomain(struct buf *bp)
1462 return (&bdomain[bp->b_domain]);
1465 static struct bufqueue *
1466 bufqueue(struct buf *bp)
1469 switch (bp->b_qindex) {
1472 case QUEUE_SENTINEL:
1477 return (&bufdomain(bp)->bd_dirtyq);
1479 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1483 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1487 * Return the locked bufqueue that bp is a member of.
1489 static struct bufqueue *
1490 bufqueue_acquire(struct buf *bp)
1492 struct bufqueue *bq, *nbq;
1495 * bp can be pushed from a per-cpu queue to the
1496 * cleanq while we're waiting on the lock. Retry
1497 * if the queues don't match.
1515 * Insert the buffer into the appropriate free list. Requires a
1516 * locked buffer on entry and buffer is unlocked before return.
1519 binsfree(struct buf *bp, int qindex)
1521 struct bufdomain *bd;
1522 struct bufqueue *bq;
1524 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1525 ("binsfree: Invalid qindex %d", qindex));
1526 BUF_ASSERT_XLOCKED(bp);
1529 * Handle delayed bremfree() processing.
1531 if (bp->b_flags & B_REMFREE) {
1532 if (bp->b_qindex == qindex) {
1533 bp->b_flags |= B_REUSE;
1534 bp->b_flags &= ~B_REMFREE;
1538 bq = bufqueue_acquire(bp);
1543 if (qindex == QUEUE_CLEAN) {
1544 if (bd->bd_lim != 0)
1545 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1549 bq = &bd->bd_dirtyq;
1550 bq_insert(bq, bp, true);
1556 * Free a buffer to the buf zone once it no longer has valid contents.
1559 buf_free(struct buf *bp)
1562 if (bp->b_flags & B_REMFREE)
1564 if (bp->b_vflags & BV_BKGRDINPROG)
1565 panic("losing buffer 1");
1566 if (bp->b_rcred != NOCRED) {
1567 crfree(bp->b_rcred);
1568 bp->b_rcred = NOCRED;
1570 if (bp->b_wcred != NOCRED) {
1571 crfree(bp->b_wcred);
1572 bp->b_wcred = NOCRED;
1574 if (!LIST_EMPTY(&bp->b_dep))
1577 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1579 uma_zfree(buf_zone, bp);
1585 * Import bufs into the uma cache from the buf list. The system still
1586 * expects a static array of bufs and much of the synchronization
1587 * around bufs assumes type stable storage. As a result, UMA is used
1588 * only as a per-cpu cache of bufs still maintained on a global list.
1591 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1597 for (i = 0; i < cnt; i++) {
1598 bp = TAILQ_FIRST(&bqempty.bq_queue);
1601 bq_remove(&bqempty, bp);
1604 BQ_UNLOCK(&bqempty);
1612 * Release bufs from the uma cache back to the buffer queues.
1615 buf_release(void *arg, void **store, int cnt)
1617 struct bufqueue *bq;
1623 for (i = 0; i < cnt; i++) {
1625 /* Inline bq_insert() to batch locking. */
1626 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1627 bp->b_flags &= ~(B_AGE | B_REUSE);
1629 bp->b_qindex = bq->bq_index;
1637 * Allocate an empty buffer header.
1640 buf_alloc(struct bufdomain *bd)
1646 * We can only run out of bufs in the buf zone if the average buf
1647 * is less than BKVASIZE. In this case the actual wait/block will
1648 * come from buf_reycle() failing to flush one of these small bufs.
1651 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1653 bp = uma_zalloc(buf_zone, M_NOWAIT);
1655 atomic_add_int(&bd->bd_freebuffers, 1);
1656 bufspace_daemon_wakeup(bd);
1657 counter_u64_add(numbufallocfails, 1);
1661 * Wake-up the bufspace daemon on transition below threshold.
1663 if (freebufs == bd->bd_lofreebuffers)
1664 bufspace_daemon_wakeup(bd);
1666 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1667 panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1669 KASSERT(bp->b_vp == NULL,
1670 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1671 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1672 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1673 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1674 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1675 KASSERT(bp->b_npages == 0,
1676 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1677 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1678 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1680 bp->b_domain = BD_DOMAIN(bd);
1686 bp->b_blkno = bp->b_lblkno = 0;
1687 bp->b_offset = NOOFFSET;
1693 bp->b_dirtyoff = bp->b_dirtyend = 0;
1694 bp->b_bufobj = NULL;
1695 bp->b_data = bp->b_kvabase = unmapped_buf;
1696 bp->b_fsprivate1 = NULL;
1697 bp->b_fsprivate2 = NULL;
1698 bp->b_fsprivate3 = NULL;
1699 LIST_INIT(&bp->b_dep);
1707 * Free a buffer from the given bufqueue. kva controls whether the
1708 * freed buf must own some kva resources. This is used for
1712 buf_recycle(struct bufdomain *bd, bool kva)
1714 struct bufqueue *bq;
1715 struct buf *bp, *nbp;
1718 counter_u64_add(bufdefragcnt, 1);
1722 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1723 ("buf_recycle: Locks don't match"));
1724 nbp = TAILQ_FIRST(&bq->bq_queue);
1727 * Run scan, possibly freeing data and/or kva mappings on the fly
1730 while ((bp = nbp) != NULL) {
1732 * Calculate next bp (we can only use it if we do not
1733 * release the bqlock).
1735 nbp = TAILQ_NEXT(bp, b_freelist);
1738 * If we are defragging then we need a buffer with
1739 * some kva to reclaim.
1741 if (kva && bp->b_kvasize == 0)
1744 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1748 * Implement a second chance algorithm for frequently
1751 if ((bp->b_flags & B_REUSE) != 0) {
1752 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1753 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1754 bp->b_flags &= ~B_REUSE;
1760 * Skip buffers with background writes in progress.
1762 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1767 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1768 ("buf_recycle: inconsistent queue %d bp %p",
1770 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1771 ("getnewbuf: queue domain %d doesn't match request %d",
1772 bp->b_domain, (int)BD_DOMAIN(bd)));
1774 * NOTE: nbp is now entirely invalid. We can only restart
1775 * the scan from this point on.
1781 * Requeue the background write buffer with error and
1784 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1787 nbp = TAILQ_FIRST(&bq->bq_queue);
1790 bp->b_flags |= B_INVAL;
1803 * Mark the buffer for removal from the appropriate free list.
1807 bremfree(struct buf *bp)
1810 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1811 KASSERT((bp->b_flags & B_REMFREE) == 0,
1812 ("bremfree: buffer %p already marked for delayed removal.", bp));
1813 KASSERT(bp->b_qindex != QUEUE_NONE,
1814 ("bremfree: buffer %p not on a queue.", bp));
1815 BUF_ASSERT_XLOCKED(bp);
1817 bp->b_flags |= B_REMFREE;
1823 * Force an immediate removal from a free list. Used only in nfs when
1824 * it abuses the b_freelist pointer.
1827 bremfreef(struct buf *bp)
1829 struct bufqueue *bq;
1831 bq = bufqueue_acquire(bp);
1837 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1840 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1841 TAILQ_INIT(&bq->bq_queue);
1843 bq->bq_index = qindex;
1844 bq->bq_subqueue = subqueue;
1848 bd_init(struct bufdomain *bd)
1852 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1853 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1854 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1855 for (i = 0; i <= mp_maxid; i++)
1856 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1857 "bufq clean subqueue lock");
1858 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1864 * Removes a buffer from the free list, must be called with the
1865 * correct qlock held.
1868 bq_remove(struct bufqueue *bq, struct buf *bp)
1871 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1872 bp, bp->b_vp, bp->b_flags);
1873 KASSERT(bp->b_qindex != QUEUE_NONE,
1874 ("bq_remove: buffer %p not on a queue.", bp));
1875 KASSERT(bufqueue(bp) == bq,
1876 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1878 BQ_ASSERT_LOCKED(bq);
1879 if (bp->b_qindex != QUEUE_EMPTY) {
1880 BUF_ASSERT_XLOCKED(bp);
1882 KASSERT(bq->bq_len >= 1,
1883 ("queue %d underflow", bp->b_qindex));
1884 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1886 bp->b_qindex = QUEUE_NONE;
1887 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1891 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1895 BQ_ASSERT_LOCKED(bq);
1896 if (bq != bd->bd_cleanq) {
1898 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1899 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1900 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1902 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1904 bd->bd_cleanq->bq_len += bq->bq_len;
1907 if (bd->bd_wanted) {
1909 wakeup(&bd->bd_wanted);
1911 if (bq != bd->bd_cleanq)
1916 bd_flushall(struct bufdomain *bd)
1918 struct bufqueue *bq;
1922 if (bd->bd_lim == 0)
1925 for (i = 0; i <= mp_maxid; i++) {
1926 bq = &bd->bd_subq[i];
1927 if (bq->bq_len == 0)
1939 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1941 struct bufdomain *bd;
1943 if (bp->b_qindex != QUEUE_NONE)
1944 panic("bq_insert: free buffer %p onto another queue?", bp);
1947 if (bp->b_flags & B_AGE) {
1948 /* Place this buf directly on the real queue. */
1949 if (bq->bq_index == QUEUE_CLEAN)
1952 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1955 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1957 bp->b_flags &= ~(B_AGE | B_REUSE);
1959 bp->b_qindex = bq->bq_index;
1960 bp->b_subqueue = bq->bq_subqueue;
1963 * Unlock before we notify so that we don't wakeup a waiter that
1964 * fails a trylock on the buf and sleeps again.
1969 if (bp->b_qindex == QUEUE_CLEAN) {
1971 * Flush the per-cpu queue and notify any waiters.
1973 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1974 bq->bq_len >= bd->bd_lim))
1983 * Free the kva allocation for a buffer.
1987 bufkva_free(struct buf *bp)
1991 if (bp->b_kvasize == 0) {
1992 KASSERT(bp->b_kvabase == unmapped_buf &&
1993 bp->b_data == unmapped_buf,
1994 ("Leaked KVA space on %p", bp));
1995 } else if (buf_mapped(bp))
1996 BUF_CHECK_MAPPED(bp);
1998 BUF_CHECK_UNMAPPED(bp);
2000 if (bp->b_kvasize == 0)
2003 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2004 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2005 counter_u64_add(buffreekvacnt, 1);
2006 bp->b_data = bp->b_kvabase = unmapped_buf;
2013 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2016 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2021 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2022 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2027 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2030 * Buffer map is too fragmented. Request the caller
2031 * to defragment the map.
2035 bp->b_kvabase = (caddr_t)addr;
2036 bp->b_kvasize = maxsize;
2037 counter_u64_add(bufkvaspace, bp->b_kvasize);
2038 if ((gbflags & GB_UNMAPPED) != 0) {
2039 bp->b_data = unmapped_buf;
2040 BUF_CHECK_UNMAPPED(bp);
2042 bp->b_data = bp->b_kvabase;
2043 BUF_CHECK_MAPPED(bp);
2051 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2052 * callback that fires to avoid returning failure.
2055 bufkva_reclaim(vmem_t *vmem, int flags)
2062 for (i = 0; i < 5; i++) {
2063 for (q = 0; q < buf_domains; q++)
2064 if (buf_recycle(&bdomain[q], true) != 0)
2073 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2074 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2075 * the buffer is valid and we do not have to do anything.
2078 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2079 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2087 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2088 if (inmem(vp, *rablkno))
2090 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2091 if ((rabp->b_flags & B_CACHE) != 0) {
2098 racct_add_buf(curproc, rabp, 0);
2099 PROC_UNLOCK(curproc);
2102 td->td_ru.ru_inblock++;
2103 rabp->b_flags |= B_ASYNC;
2104 rabp->b_flags &= ~B_INVAL;
2105 if ((flags & GB_CKHASH) != 0) {
2106 rabp->b_flags |= B_CKHASH;
2107 rabp->b_ckhashcalc = ckhashfunc;
2109 rabp->b_ioflags &= ~BIO_ERROR;
2110 rabp->b_iocmd = BIO_READ;
2111 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2112 rabp->b_rcred = crhold(cred);
2113 vfs_busy_pages(rabp, 0);
2115 rabp->b_iooffset = dbtob(rabp->b_blkno);
2121 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2123 * Get a buffer with the specified data. Look in the cache first. We
2124 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2125 * is set, the buffer is valid and we do not have to do anything, see
2126 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2128 * Always return a NULL buffer pointer (in bpp) when returning an error.
2131 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2132 int *rabsize, int cnt, struct ucred *cred, int flags,
2133 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2137 int error, readwait, rv;
2139 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2142 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2145 error = getblkx(vp, blkno, size, 0, 0, flags, &bp);
2150 flags &= ~GB_NOSPARSE;
2154 * If not found in cache, do some I/O
2157 if ((bp->b_flags & B_CACHE) == 0) {
2160 PROC_LOCK(td->td_proc);
2161 racct_add_buf(td->td_proc, bp, 0);
2162 PROC_UNLOCK(td->td_proc);
2165 td->td_ru.ru_inblock++;
2166 bp->b_iocmd = BIO_READ;
2167 bp->b_flags &= ~B_INVAL;
2168 if ((flags & GB_CKHASH) != 0) {
2169 bp->b_flags |= B_CKHASH;
2170 bp->b_ckhashcalc = ckhashfunc;
2172 bp->b_ioflags &= ~BIO_ERROR;
2173 if (bp->b_rcred == NOCRED && cred != NOCRED)
2174 bp->b_rcred = crhold(cred);
2175 vfs_busy_pages(bp, 0);
2176 bp->b_iooffset = dbtob(bp->b_blkno);
2182 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2184 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2198 * Write, release buffer on completion. (Done by iodone
2199 * if async). Do not bother writing anything if the buffer
2202 * Note that we set B_CACHE here, indicating that buffer is
2203 * fully valid and thus cacheable. This is true even of NFS
2204 * now so we set it generally. This could be set either here
2205 * or in biodone() since the I/O is synchronous. We put it
2209 bufwrite(struct buf *bp)
2216 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2217 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2218 bp->b_flags |= B_INVAL | B_RELBUF;
2219 bp->b_flags &= ~B_CACHE;
2223 if (bp->b_flags & B_INVAL) {
2228 if (bp->b_flags & B_BARRIER)
2229 atomic_add_long(&barrierwrites, 1);
2231 oldflags = bp->b_flags;
2233 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2234 ("FFS background buffer should not get here %p", bp));
2238 vp_md = vp->v_vflag & VV_MD;
2243 * Mark the buffer clean. Increment the bufobj write count
2244 * before bundirty() call, to prevent other thread from seeing
2245 * empty dirty list and zero counter for writes in progress,
2246 * falsely indicating that the bufobj is clean.
2248 bufobj_wref(bp->b_bufobj);
2251 bp->b_flags &= ~B_DONE;
2252 bp->b_ioflags &= ~BIO_ERROR;
2253 bp->b_flags |= B_CACHE;
2254 bp->b_iocmd = BIO_WRITE;
2256 vfs_busy_pages(bp, 1);
2259 * Normal bwrites pipeline writes
2261 bp->b_runningbufspace = bp->b_bufsize;
2262 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2267 racct_add_buf(curproc, bp, 1);
2268 PROC_UNLOCK(curproc);
2271 curthread->td_ru.ru_oublock++;
2272 if (oldflags & B_ASYNC)
2274 bp->b_iooffset = dbtob(bp->b_blkno);
2275 buf_track(bp, __func__);
2278 if ((oldflags & B_ASYNC) == 0) {
2279 int rtval = bufwait(bp);
2282 } else if (space > hirunningspace) {
2284 * don't allow the async write to saturate the I/O
2285 * system. We will not deadlock here because
2286 * we are blocking waiting for I/O that is already in-progress
2287 * to complete. We do not block here if it is the update
2288 * or syncer daemon trying to clean up as that can lead
2291 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2292 waitrunningbufspace();
2299 bufbdflush(struct bufobj *bo, struct buf *bp)
2303 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2304 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2306 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2309 * Try to find a buffer to flush.
2311 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2312 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2314 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2317 panic("bdwrite: found ourselves");
2319 /* Don't countdeps with the bo lock held. */
2320 if (buf_countdeps(nbp, 0)) {
2325 if (nbp->b_flags & B_CLUSTEROK) {
2326 vfs_bio_awrite(nbp);
2331 dirtybufferflushes++;
2340 * Delayed write. (Buffer is marked dirty). Do not bother writing
2341 * anything if the buffer is marked invalid.
2343 * Note that since the buffer must be completely valid, we can safely
2344 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2345 * biodone() in order to prevent getblk from writing the buffer
2346 * out synchronously.
2349 bdwrite(struct buf *bp)
2351 struct thread *td = curthread;
2355 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2356 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2357 KASSERT((bp->b_flags & B_BARRIER) == 0,
2358 ("Barrier request in delayed write %p", bp));
2360 if (bp->b_flags & B_INVAL) {
2366 * If we have too many dirty buffers, don't create any more.
2367 * If we are wildly over our limit, then force a complete
2368 * cleanup. Otherwise, just keep the situation from getting
2369 * out of control. Note that we have to avoid a recursive
2370 * disaster and not try to clean up after our own cleanup!
2374 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2375 td->td_pflags |= TDP_INBDFLUSH;
2377 td->td_pflags &= ~TDP_INBDFLUSH;
2383 * Set B_CACHE, indicating that the buffer is fully valid. This is
2384 * true even of NFS now.
2386 bp->b_flags |= B_CACHE;
2389 * This bmap keeps the system from needing to do the bmap later,
2390 * perhaps when the system is attempting to do a sync. Since it
2391 * is likely that the indirect block -- or whatever other datastructure
2392 * that the filesystem needs is still in memory now, it is a good
2393 * thing to do this. Note also, that if the pageout daemon is
2394 * requesting a sync -- there might not be enough memory to do
2395 * the bmap then... So, this is important to do.
2397 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2398 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2401 buf_track(bp, __func__);
2404 * Set the *dirty* buffer range based upon the VM system dirty
2407 * Mark the buffer pages as clean. We need to do this here to
2408 * satisfy the vnode_pager and the pageout daemon, so that it
2409 * thinks that the pages have been "cleaned". Note that since
2410 * the pages are in a delayed write buffer -- the VFS layer
2411 * "will" see that the pages get written out on the next sync,
2412 * or perhaps the cluster will be completed.
2414 vfs_clean_pages_dirty_buf(bp);
2418 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2419 * due to the softdep code.
2426 * Turn buffer into delayed write request. We must clear BIO_READ and
2427 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2428 * itself to properly update it in the dirty/clean lists. We mark it
2429 * B_DONE to ensure that any asynchronization of the buffer properly
2430 * clears B_DONE ( else a panic will occur later ).
2432 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2433 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2434 * should only be called if the buffer is known-good.
2436 * Since the buffer is not on a queue, we do not update the numfreebuffers
2439 * The buffer must be on QUEUE_NONE.
2442 bdirty(struct buf *bp)
2445 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2446 bp, bp->b_vp, bp->b_flags);
2447 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2448 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2449 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2450 bp->b_flags &= ~(B_RELBUF);
2451 bp->b_iocmd = BIO_WRITE;
2453 if ((bp->b_flags & B_DELWRI) == 0) {
2454 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2463 * Clear B_DELWRI for buffer.
2465 * Since the buffer is not on a queue, we do not update the numfreebuffers
2468 * The buffer must be on QUEUE_NONE.
2472 bundirty(struct buf *bp)
2475 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2476 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2477 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2478 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2480 if (bp->b_flags & B_DELWRI) {
2481 bp->b_flags &= ~B_DELWRI;
2486 * Since it is now being written, we can clear its deferred write flag.
2488 bp->b_flags &= ~B_DEFERRED;
2494 * Asynchronous write. Start output on a buffer, but do not wait for
2495 * it to complete. The buffer is released when the output completes.
2497 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2498 * B_INVAL buffers. Not us.
2501 bawrite(struct buf *bp)
2504 bp->b_flags |= B_ASYNC;
2511 * Asynchronous barrier write. Start output on a buffer, but do not
2512 * wait for it to complete. Place a write barrier after this write so
2513 * that this buffer and all buffers written before it are committed to
2514 * the disk before any buffers written after this write are committed
2515 * to the disk. The buffer is released when the output completes.
2518 babarrierwrite(struct buf *bp)
2521 bp->b_flags |= B_ASYNC | B_BARRIER;
2528 * Synchronous barrier write. Start output on a buffer and wait for
2529 * it to complete. Place a write barrier after this write so that
2530 * this buffer and all buffers written before it are committed to
2531 * the disk before any buffers written after this write are committed
2532 * to the disk. The buffer is released when the output completes.
2535 bbarrierwrite(struct buf *bp)
2538 bp->b_flags |= B_BARRIER;
2539 return (bwrite(bp));
2545 * Called prior to the locking of any vnodes when we are expecting to
2546 * write. We do not want to starve the buffer cache with too many
2547 * dirty buffers so we block here. By blocking prior to the locking
2548 * of any vnodes we attempt to avoid the situation where a locked vnode
2549 * prevents the various system daemons from flushing related buffers.
2555 if (buf_dirty_count_severe()) {
2556 mtx_lock(&bdirtylock);
2557 while (buf_dirty_count_severe()) {
2559 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2562 mtx_unlock(&bdirtylock);
2567 * Return true if we have too many dirty buffers.
2570 buf_dirty_count_severe(void)
2573 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2579 * Release a busy buffer and, if requested, free its resources. The
2580 * buffer will be stashed in the appropriate bufqueue[] allowing it
2581 * to be accessed later as a cache entity or reused for other purposes.
2584 brelse(struct buf *bp)
2586 struct mount *v_mnt;
2590 * Many functions erroneously call brelse with a NULL bp under rare
2591 * error conditions. Simply return when called with a NULL bp.
2595 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2596 bp, bp->b_vp, bp->b_flags);
2597 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2598 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2599 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2600 ("brelse: non-VMIO buffer marked NOREUSE"));
2602 if (BUF_LOCKRECURSED(bp)) {
2604 * Do not process, in particular, do not handle the
2605 * B_INVAL/B_RELBUF and do not release to free list.
2611 if (bp->b_flags & B_MANAGED) {
2616 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2617 BO_LOCK(bp->b_bufobj);
2618 bp->b_vflags &= ~BV_BKGRDERR;
2619 BO_UNLOCK(bp->b_bufobj);
2623 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2624 (bp->b_flags & B_INVALONERR)) {
2626 * Forced invalidation of dirty buffer contents, to be used
2627 * after a failed write in the rare case that the loss of the
2628 * contents is acceptable. The buffer is invalidated and
2631 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2632 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2635 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2636 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2637 !(bp->b_flags & B_INVAL)) {
2639 * Failed write, redirty. All errors except ENXIO (which
2640 * means the device is gone) are treated as being
2643 * XXX Treating EIO as transient is not correct; the
2644 * contract with the local storage device drivers is that
2645 * they will only return EIO once the I/O is no longer
2646 * retriable. Network I/O also respects this through the
2647 * guarantees of TCP and/or the internal retries of NFS.
2648 * ENOMEM might be transient, but we also have no way of
2649 * knowing when its ok to retry/reschedule. In general,
2650 * this entire case should be made obsolete through better
2651 * error handling/recovery and resource scheduling.
2653 * Do this also for buffers that failed with ENXIO, but have
2654 * non-empty dependencies - the soft updates code might need
2655 * to access the buffer to untangle them.
2657 * Must clear BIO_ERROR to prevent pages from being scrapped.
2659 bp->b_ioflags &= ~BIO_ERROR;
2661 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2662 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2664 * Either a failed read I/O, or we were asked to free or not
2665 * cache the buffer, or we failed to write to a device that's
2666 * no longer present.
2668 bp->b_flags |= B_INVAL;
2669 if (!LIST_EMPTY(&bp->b_dep))
2671 if (bp->b_flags & B_DELWRI)
2673 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2674 if ((bp->b_flags & B_VMIO) == 0) {
2682 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2683 * is called with B_DELWRI set, the underlying pages may wind up
2684 * getting freed causing a previous write (bdwrite()) to get 'lost'
2685 * because pages associated with a B_DELWRI bp are marked clean.
2687 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2688 * if B_DELWRI is set.
2690 if (bp->b_flags & B_DELWRI)
2691 bp->b_flags &= ~B_RELBUF;
2694 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2695 * constituted, not even NFS buffers now. Two flags effect this. If
2696 * B_INVAL, the struct buf is invalidated but the VM object is kept
2697 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2699 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2700 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2701 * buffer is also B_INVAL because it hits the re-dirtying code above.
2703 * Normally we can do this whether a buffer is B_DELWRI or not. If
2704 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2705 * the commit state and we cannot afford to lose the buffer. If the
2706 * buffer has a background write in progress, we need to keep it
2707 * around to prevent it from being reconstituted and starting a second
2711 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2713 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2714 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2715 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2716 vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2717 vfs_vmio_invalidate(bp);
2721 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2722 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2724 bp->b_flags &= ~B_NOREUSE;
2725 if (bp->b_vp != NULL)
2730 * If the buffer has junk contents signal it and eventually
2731 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2734 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2735 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2736 bp->b_flags |= B_INVAL;
2737 if (bp->b_flags & B_INVAL) {
2738 if (bp->b_flags & B_DELWRI)
2744 buf_track(bp, __func__);
2746 /* buffers with no memory */
2747 if (bp->b_bufsize == 0) {
2751 /* buffers with junk contents */
2752 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2753 (bp->b_ioflags & BIO_ERROR)) {
2754 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2755 if (bp->b_vflags & BV_BKGRDINPROG)
2756 panic("losing buffer 2");
2757 qindex = QUEUE_CLEAN;
2758 bp->b_flags |= B_AGE;
2759 /* remaining buffers */
2760 } else if (bp->b_flags & B_DELWRI)
2761 qindex = QUEUE_DIRTY;
2763 qindex = QUEUE_CLEAN;
2765 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2766 panic("brelse: not dirty");
2768 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2769 /* binsfree unlocks bp. */
2770 binsfree(bp, qindex);
2774 * Release a buffer back to the appropriate queue but do not try to free
2775 * it. The buffer is expected to be used again soon.
2777 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2778 * biodone() to requeue an async I/O on completion. It is also used when
2779 * known good buffers need to be requeued but we think we may need the data
2782 * XXX we should be able to leave the B_RELBUF hint set on completion.
2785 bqrelse(struct buf *bp)
2789 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2790 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2791 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2793 qindex = QUEUE_NONE;
2794 if (BUF_LOCKRECURSED(bp)) {
2795 /* do not release to free list */
2799 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2801 if (bp->b_flags & B_MANAGED) {
2802 if (bp->b_flags & B_REMFREE)
2807 /* buffers with stale but valid contents */
2808 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2809 BV_BKGRDERR)) == BV_BKGRDERR) {
2810 BO_LOCK(bp->b_bufobj);
2811 bp->b_vflags &= ~BV_BKGRDERR;
2812 BO_UNLOCK(bp->b_bufobj);
2813 qindex = QUEUE_DIRTY;
2815 if ((bp->b_flags & B_DELWRI) == 0 &&
2816 (bp->b_xflags & BX_VNDIRTY))
2817 panic("bqrelse: not dirty");
2818 if ((bp->b_flags & B_NOREUSE) != 0) {
2822 qindex = QUEUE_CLEAN;
2824 buf_track(bp, __func__);
2825 /* binsfree unlocks bp. */
2826 binsfree(bp, qindex);
2830 buf_track(bp, __func__);
2836 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2837 * restore bogus pages.
2840 vfs_vmio_iodone(struct buf *bp)
2845 struct vnode *vp __unused;
2846 int i, iosize, resid;
2849 obj = bp->b_bufobj->bo_object;
2850 KASSERT(REFCOUNT_COUNT(obj->paging_in_progress) >= bp->b_npages,
2851 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2852 REFCOUNT_COUNT(obj->paging_in_progress), bp->b_npages));
2855 KASSERT(vp->v_holdcnt > 0,
2856 ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2857 KASSERT(vp->v_object != NULL,
2858 ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2860 foff = bp->b_offset;
2861 KASSERT(bp->b_offset != NOOFFSET,
2862 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2865 iosize = bp->b_bcount - bp->b_resid;
2866 for (i = 0; i < bp->b_npages; i++) {
2867 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2872 * cleanup bogus pages, restoring the originals
2875 if (m == bogus_page) {
2876 if (bogus == false) {
2878 VM_OBJECT_RLOCK(obj);
2880 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2882 panic("biodone: page disappeared!");
2884 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2886 * In the write case, the valid and clean bits are
2887 * already changed correctly ( see bdwrite() ), so we
2888 * only need to do this here in the read case.
2890 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2891 resid)) == 0, ("vfs_vmio_iodone: page %p "
2892 "has unexpected dirty bits", m));
2893 vfs_page_set_valid(bp, foff, m);
2895 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2896 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2897 (intmax_t)foff, (uintmax_t)m->pindex));
2900 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2904 VM_OBJECT_RUNLOCK(obj);
2905 vm_object_pip_wakeupn(obj, bp->b_npages);
2906 if (bogus && buf_mapped(bp)) {
2907 BUF_CHECK_MAPPED(bp);
2908 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2909 bp->b_pages, bp->b_npages);
2914 * Perform page invalidation when a buffer is released. The fully invalid
2915 * pages will be reclaimed later in vfs_vmio_truncate().
2918 vfs_vmio_invalidate(struct buf *bp)
2922 int flags, i, resid, poffset, presid;
2924 if (buf_mapped(bp)) {
2925 BUF_CHECK_MAPPED(bp);
2926 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2928 BUF_CHECK_UNMAPPED(bp);
2930 * Get the base offset and length of the buffer. Note that
2931 * in the VMIO case if the buffer block size is not
2932 * page-aligned then b_data pointer may not be page-aligned.
2933 * But our b_pages[] array *IS* page aligned.
2935 * block sizes less then DEV_BSIZE (usually 512) are not
2936 * supported due to the page granularity bits (m->valid,
2937 * m->dirty, etc...).
2939 * See man buf(9) for more information
2941 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2942 obj = bp->b_bufobj->bo_object;
2943 resid = bp->b_bufsize;
2944 poffset = bp->b_offset & PAGE_MASK;
2945 VM_OBJECT_WLOCK(obj);
2946 for (i = 0; i < bp->b_npages; i++) {
2948 if (m == bogus_page)
2949 panic("vfs_vmio_invalidate: Unexpected bogus page.");
2950 bp->b_pages[i] = NULL;
2952 presid = resid > (PAGE_SIZE - poffset) ?
2953 (PAGE_SIZE - poffset) : resid;
2954 KASSERT(presid >= 0, ("brelse: extra page"));
2955 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2956 if (pmap_page_wired_mappings(m) == 0)
2957 vm_page_set_invalid(m, poffset, presid);
2959 vm_page_release_locked(m, flags);
2963 VM_OBJECT_WUNLOCK(obj);
2968 * Page-granular truncation of an existing VMIO buffer.
2971 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2977 if (bp->b_npages == desiredpages)
2980 if (buf_mapped(bp)) {
2981 BUF_CHECK_MAPPED(bp);
2982 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2983 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2985 BUF_CHECK_UNMAPPED(bp);
2988 * The object lock is needed only if we will attempt to free pages.
2990 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2991 if ((bp->b_flags & B_DIRECT) != 0) {
2992 flags |= VPR_TRYFREE;
2993 obj = bp->b_bufobj->bo_object;
2994 VM_OBJECT_WLOCK(obj);
2998 for (i = desiredpages; i < bp->b_npages; i++) {
3000 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3001 bp->b_pages[i] = NULL;
3003 vm_page_release_locked(m, flags);
3005 vm_page_release(m, flags);
3008 VM_OBJECT_WUNLOCK(obj);
3009 bp->b_npages = desiredpages;
3013 * Byte granular extension of VMIO buffers.
3016 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3019 * We are growing the buffer, possibly in a
3020 * byte-granular fashion.
3028 * Step 1, bring in the VM pages from the object, allocating
3029 * them if necessary. We must clear B_CACHE if these pages
3030 * are not valid for the range covered by the buffer.
3032 obj = bp->b_bufobj->bo_object;
3033 if (bp->b_npages < desiredpages) {
3035 * We must allocate system pages since blocking
3036 * here could interfere with paging I/O, no
3037 * matter which process we are.
3039 * Only exclusive busy can be tested here.
3040 * Blocking on shared busy might lead to
3041 * deadlocks once allocbuf() is called after
3042 * pages are vfs_busy_pages().
3044 VM_OBJECT_WLOCK(obj);
3045 (void)vm_page_grab_pages(obj,
3046 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3047 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3048 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3049 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3050 VM_OBJECT_WUNLOCK(obj);
3051 bp->b_npages = desiredpages;
3055 * Step 2. We've loaded the pages into the buffer,
3056 * we have to figure out if we can still have B_CACHE
3057 * set. Note that B_CACHE is set according to the
3058 * byte-granular range ( bcount and size ), not the
3059 * aligned range ( newbsize ).
3061 * The VM test is against m->valid, which is DEV_BSIZE
3062 * aligned. Needless to say, the validity of the data
3063 * needs to also be DEV_BSIZE aligned. Note that this
3064 * fails with NFS if the server or some other client
3065 * extends the file's EOF. If our buffer is resized,
3066 * B_CACHE may remain set! XXX
3068 toff = bp->b_bcount;
3069 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3070 while ((bp->b_flags & B_CACHE) && toff < size) {
3073 if (tinc > (size - toff))
3075 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3076 m = bp->b_pages[pi];
3077 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3083 * Step 3, fixup the KVA pmap.
3088 BUF_CHECK_UNMAPPED(bp);
3092 * Check to see if a block at a particular lbn is available for a clustered
3096 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3103 /* If the buf isn't in core skip it */
3104 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3107 /* If the buf is busy we don't want to wait for it */
3108 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3111 /* Only cluster with valid clusterable delayed write buffers */
3112 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3113 (B_DELWRI | B_CLUSTEROK))
3116 if (bpa->b_bufsize != size)
3120 * Check to see if it is in the expected place on disk and that the
3121 * block has been mapped.
3123 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3133 * Implement clustered async writes for clearing out B_DELWRI buffers.
3134 * This is much better then the old way of writing only one buffer at
3135 * a time. Note that we may not be presented with the buffers in the
3136 * correct order, so we search for the cluster in both directions.
3139 vfs_bio_awrite(struct buf *bp)
3144 daddr_t lblkno = bp->b_lblkno;
3145 struct vnode *vp = bp->b_vp;
3153 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3155 * right now we support clustered writing only to regular files. If
3156 * we find a clusterable block we could be in the middle of a cluster
3157 * rather then at the beginning.
3159 if ((vp->v_type == VREG) &&
3160 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3161 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3163 size = vp->v_mount->mnt_stat.f_iosize;
3164 maxcl = MAXPHYS / size;
3167 for (i = 1; i < maxcl; i++)
3168 if (vfs_bio_clcheck(vp, size, lblkno + i,
3169 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3172 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3173 if (vfs_bio_clcheck(vp, size, lblkno - j,
3174 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3180 * this is a possible cluster write
3184 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3190 bp->b_flags |= B_ASYNC;
3192 * default (old) behavior, writing out only one block
3194 * XXX returns b_bufsize instead of b_bcount for nwritten?
3196 nwritten = bp->b_bufsize;
3205 * Allocate KVA for an empty buf header according to gbflags.
3208 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3211 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3213 * In order to keep fragmentation sane we only allocate kva
3214 * in BKVASIZE chunks. XXX with vmem we can do page size.
3216 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3218 if (maxsize != bp->b_kvasize &&
3219 bufkva_alloc(bp, maxsize, gbflags))
3228 * Find and initialize a new buffer header, freeing up existing buffers
3229 * in the bufqueues as necessary. The new buffer is returned locked.
3232 * We have insufficient buffer headers
3233 * We have insufficient buffer space
3234 * buffer_arena is too fragmented ( space reservation fails )
3235 * If we have to flush dirty buffers ( but we try to avoid this )
3237 * The caller is responsible for releasing the reserved bufspace after
3238 * allocbuf() is called.
3241 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3243 struct bufdomain *bd;
3245 bool metadata, reserved;
3248 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3249 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3250 if (!unmapped_buf_allowed)
3251 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3253 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3261 bd = &bdomain[vp->v_bufobj.bo_domain];
3263 counter_u64_add(getnewbufcalls, 1);
3266 if (reserved == false &&
3267 bufspace_reserve(bd, maxsize, metadata) != 0) {
3268 counter_u64_add(getnewbufrestarts, 1);
3272 if ((bp = buf_alloc(bd)) == NULL) {
3273 counter_u64_add(getnewbufrestarts, 1);
3276 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3279 } while (buf_recycle(bd, false) == 0);
3282 bufspace_release(bd, maxsize);
3284 bp->b_flags |= B_INVAL;
3287 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3295 * buffer flushing daemon. Buffers are normally flushed by the
3296 * update daemon but if it cannot keep up this process starts to
3297 * take the load in an attempt to prevent getnewbuf() from blocking.
3299 static struct kproc_desc buf_kp = {
3304 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3307 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3311 flushed = flushbufqueues(vp, bd, target, 0);
3314 * Could not find any buffers without rollback
3315 * dependencies, so just write the first one
3316 * in the hopes of eventually making progress.
3318 if (vp != NULL && target > 2)
3320 flushbufqueues(vp, bd, target, 1);
3328 struct bufdomain *bd;
3334 * This process needs to be suspended prior to shutdown sync.
3336 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3337 SHUTDOWN_PRI_LAST + 100);
3340 * Start the buf clean daemons as children threads.
3342 for (i = 0 ; i < buf_domains; i++) {
3345 error = kthread_add((void (*)(void *))bufspace_daemon,
3346 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3348 panic("error %d spawning bufspace daemon", error);
3352 * This process is allowed to take the buffer cache to the limit
3354 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3358 mtx_unlock(&bdlock);
3360 kthread_suspend_check();
3363 * Save speedupreq for this pass and reset to capture new
3366 speedupreq = bd_speedupreq;
3370 * Flush each domain sequentially according to its level and
3371 * the speedup request.
3373 for (i = 0; i < buf_domains; i++) {
3376 lodirty = bd->bd_numdirtybuffers / 2;
3378 lodirty = bd->bd_lodirtybuffers;
3379 while (bd->bd_numdirtybuffers > lodirty) {
3380 if (buf_flush(NULL, bd,
3381 bd->bd_numdirtybuffers - lodirty) == 0)
3383 kern_yield(PRI_USER);
3388 * Only clear bd_request if we have reached our low water
3389 * mark. The buf_daemon normally waits 1 second and
3390 * then incrementally flushes any dirty buffers that have
3391 * built up, within reason.
3393 * If we were unable to hit our low water mark and couldn't
3394 * find any flushable buffers, we sleep for a short period
3395 * to avoid endless loops on unlockable buffers.
3398 if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3400 * We reached our low water mark, reset the
3401 * request and sleep until we are needed again.
3402 * The sleep is just so the suspend code works.
3406 * Do an extra wakeup in case dirty threshold
3407 * changed via sysctl and the explicit transition
3408 * out of shortfall was missed.
3411 if (runningbufspace <= lorunningspace)
3413 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3416 * We couldn't find any flushable dirty buffers but
3417 * still have too many dirty buffers, we
3418 * have to sleep and try again. (rare)
3420 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3428 * Try to flush a buffer in the dirty queue. We must be careful to
3429 * free up B_INVAL buffers instead of write them, which NFS is
3430 * particularly sensitive to.
3432 static int flushwithdeps = 0;
3433 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3435 "Number of buffers flushed with dependecies that require rollbacks");
3438 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3441 struct bufqueue *bq;
3442 struct buf *sentinel;
3452 bq = &bd->bd_dirtyq;
3454 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3455 sentinel->b_qindex = QUEUE_SENTINEL;
3457 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3459 while (flushed != target) {
3462 bp = TAILQ_NEXT(sentinel, b_freelist);
3464 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3465 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3472 * Skip sentinels inserted by other invocations of the
3473 * flushbufqueues(), taking care to not reorder them.
3475 * Only flush the buffers that belong to the
3476 * vnode locked by the curthread.
3478 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3483 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3489 * BKGRDINPROG can only be set with the buf and bufobj
3490 * locks both held. We tolerate a race to clear it here.
3492 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3493 (bp->b_flags & B_DELWRI) == 0) {
3497 if (bp->b_flags & B_INVAL) {
3504 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3505 if (flushdeps == 0) {
3513 * We must hold the lock on a vnode before writing
3514 * one of its buffers. Otherwise we may confuse, or
3515 * in the case of a snapshot vnode, deadlock the
3518 * The lock order here is the reverse of the normal
3519 * of vnode followed by buf lock. This is ok because
3520 * the NOWAIT will prevent deadlock.
3523 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3529 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3531 ASSERT_VOP_LOCKED(vp, "getbuf");
3533 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3534 vn_lock(vp, LK_TRYUPGRADE);
3537 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3538 bp, bp->b_vp, bp->b_flags);
3539 if (curproc == bufdaemonproc) {
3544 counter_u64_add(notbufdflushes, 1);
3546 vn_finished_write(mp);
3549 flushwithdeps += hasdeps;
3553 * Sleeping on runningbufspace while holding
3554 * vnode lock leads to deadlock.
3556 if (curproc == bufdaemonproc &&
3557 runningbufspace > hirunningspace)
3558 waitrunningbufspace();
3561 vn_finished_write(mp);
3565 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3567 free(sentinel, M_TEMP);
3572 * Check to see if a block is currently memory resident.
3575 incore(struct bufobj *bo, daddr_t blkno)
3580 bp = gbincore(bo, blkno);
3586 * Returns true if no I/O is needed to access the
3587 * associated VM object. This is like incore except
3588 * it also hunts around in the VM system for the data.
3592 inmem(struct vnode * vp, daddr_t blkno)
3595 vm_offset_t toff, tinc, size;
3599 ASSERT_VOP_LOCKED(vp, "inmem");
3601 if (incore(&vp->v_bufobj, blkno))
3603 if (vp->v_mount == NULL)
3610 if (size > vp->v_mount->mnt_stat.f_iosize)
3611 size = vp->v_mount->mnt_stat.f_iosize;
3612 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3614 VM_OBJECT_RLOCK(obj);
3615 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3616 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3620 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3621 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3622 if (vm_page_is_valid(m,
3623 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3626 VM_OBJECT_RUNLOCK(obj);
3630 VM_OBJECT_RUNLOCK(obj);
3635 * Set the dirty range for a buffer based on the status of the dirty
3636 * bits in the pages comprising the buffer. The range is limited
3637 * to the size of the buffer.
3639 * Tell the VM system that the pages associated with this buffer
3640 * are clean. This is used for delayed writes where the data is
3641 * going to go to disk eventually without additional VM intevention.
3643 * Note that while we only really need to clean through to b_bcount, we
3644 * just go ahead and clean through to b_bufsize.
3647 vfs_clean_pages_dirty_buf(struct buf *bp)
3649 vm_ooffset_t foff, noff, eoff;
3653 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3656 foff = bp->b_offset;
3657 KASSERT(bp->b_offset != NOOFFSET,
3658 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3660 vfs_busy_pages_acquire(bp);
3661 vfs_setdirty_range(bp);
3662 for (i = 0; i < bp->b_npages; i++) {
3663 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3665 if (eoff > bp->b_offset + bp->b_bufsize)
3666 eoff = bp->b_offset + bp->b_bufsize;
3668 vfs_page_set_validclean(bp, foff, m);
3669 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3672 vfs_busy_pages_release(bp);
3676 vfs_setdirty_range(struct buf *bp)
3678 vm_offset_t boffset;
3679 vm_offset_t eoffset;
3683 * test the pages to see if they have been modified directly
3684 * by users through the VM system.
3686 for (i = 0; i < bp->b_npages; i++)
3687 vm_page_test_dirty(bp->b_pages[i]);
3690 * Calculate the encompassing dirty range, boffset and eoffset,
3691 * (eoffset - boffset) bytes.
3694 for (i = 0; i < bp->b_npages; i++) {
3695 if (bp->b_pages[i]->dirty)
3698 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3700 for (i = bp->b_npages - 1; i >= 0; --i) {
3701 if (bp->b_pages[i]->dirty) {
3705 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3708 * Fit it to the buffer.
3711 if (eoffset > bp->b_bcount)
3712 eoffset = bp->b_bcount;
3715 * If we have a good dirty range, merge with the existing
3719 if (boffset < eoffset) {
3720 if (bp->b_dirtyoff > boffset)
3721 bp->b_dirtyoff = boffset;
3722 if (bp->b_dirtyend < eoffset)
3723 bp->b_dirtyend = eoffset;
3728 * Allocate the KVA mapping for an existing buffer.
3729 * If an unmapped buffer is provided but a mapped buffer is requested, take
3730 * also care to properly setup mappings between pages and KVA.
3733 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3735 int bsize, maxsize, need_mapping, need_kva;
3738 need_mapping = bp->b_data == unmapped_buf &&
3739 (gbflags & GB_UNMAPPED) == 0;
3740 need_kva = bp->b_kvabase == unmapped_buf &&
3741 bp->b_data == unmapped_buf &&
3742 (gbflags & GB_KVAALLOC) != 0;
3743 if (!need_mapping && !need_kva)
3746 BUF_CHECK_UNMAPPED(bp);
3748 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3750 * Buffer is not mapped, but the KVA was already
3751 * reserved at the time of the instantiation. Use the
3758 * Calculate the amount of the address space we would reserve
3759 * if the buffer was mapped.
3761 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3762 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3763 offset = blkno * bsize;
3764 maxsize = size + (offset & PAGE_MASK);
3765 maxsize = imax(maxsize, bsize);
3767 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3768 if ((gbflags & GB_NOWAIT_BD) != 0) {
3770 * XXXKIB: defragmentation cannot
3771 * succeed, not sure what else to do.
3773 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3775 counter_u64_add(mappingrestarts, 1);
3776 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3780 /* b_offset is handled by bpmap_qenter. */
3781 bp->b_data = bp->b_kvabase;
3782 BUF_CHECK_MAPPED(bp);
3788 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3794 error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp);
3803 * Get a block given a specified block and offset into a file/device.
3804 * The buffers B_DONE bit will be cleared on return, making it almost
3805 * ready for an I/O initiation. B_INVAL may or may not be set on
3806 * return. The caller should clear B_INVAL prior to initiating a
3809 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3810 * an existing buffer.
3812 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3813 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3814 * and then cleared based on the backing VM. If the previous buffer is
3815 * non-0-sized but invalid, B_CACHE will be cleared.
3817 * If getblk() must create a new buffer, the new buffer is returned with
3818 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3819 * case it is returned with B_INVAL clear and B_CACHE set based on the
3822 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
3823 * B_CACHE bit is clear.
3825 * What this means, basically, is that the caller should use B_CACHE to
3826 * determine whether the buffer is fully valid or not and should clear
3827 * B_INVAL prior to issuing a read. If the caller intends to validate
3828 * the buffer by loading its data area with something, the caller needs
3829 * to clear B_INVAL. If the caller does this without issuing an I/O,
3830 * the caller should set B_CACHE ( as an optimization ), else the caller
3831 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3832 * a write attempt or if it was a successful read. If the caller
3833 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3834 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3837 getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3838 int flags, struct buf **bpp)
3843 int bsize, error, maxsize, vmio;
3846 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3847 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3848 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3849 ASSERT_VOP_LOCKED(vp, "getblk");
3850 if (size > maxbcachebuf)
3851 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3853 if (!unmapped_buf_allowed)
3854 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3860 bp = gbincore(bo, blkno);
3864 * Buffer is in-core. If the buffer is not busy nor managed,
3865 * it must be on a queue.
3867 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3869 if ((flags & GB_LOCK_NOWAIT) != 0)
3870 lockflags |= LK_NOWAIT;
3872 error = BUF_TIMELOCK(bp, lockflags,
3873 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3876 * If we slept and got the lock we have to restart in case
3877 * the buffer changed identities.
3879 if (error == ENOLCK)
3881 /* We timed out or were interrupted. */
3882 else if (error != 0)
3884 /* If recursed, assume caller knows the rules. */
3885 else if (BUF_LOCKRECURSED(bp))
3889 * The buffer is locked. B_CACHE is cleared if the buffer is
3890 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3891 * and for a VMIO buffer B_CACHE is adjusted according to the
3894 if (bp->b_flags & B_INVAL)
3895 bp->b_flags &= ~B_CACHE;
3896 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3897 bp->b_flags |= B_CACHE;
3898 if (bp->b_flags & B_MANAGED)
3899 MPASS(bp->b_qindex == QUEUE_NONE);
3904 * check for size inconsistencies for non-VMIO case.
3906 if (bp->b_bcount != size) {
3907 if ((bp->b_flags & B_VMIO) == 0 ||
3908 (size > bp->b_kvasize)) {
3909 if (bp->b_flags & B_DELWRI) {
3910 bp->b_flags |= B_NOCACHE;
3913 if (LIST_EMPTY(&bp->b_dep)) {
3914 bp->b_flags |= B_RELBUF;
3917 bp->b_flags |= B_NOCACHE;
3926 * Handle the case of unmapped buffer which should
3927 * become mapped, or the buffer for which KVA
3928 * reservation is requested.
3930 bp_unmapped_get_kva(bp, blkno, size, flags);
3933 * If the size is inconsistent in the VMIO case, we can resize
3934 * the buffer. This might lead to B_CACHE getting set or
3935 * cleared. If the size has not changed, B_CACHE remains
3936 * unchanged from its previous state.
3940 KASSERT(bp->b_offset != NOOFFSET,
3941 ("getblk: no buffer offset"));
3944 * A buffer with B_DELWRI set and B_CACHE clear must
3945 * be committed before we can return the buffer in
3946 * order to prevent the caller from issuing a read
3947 * ( due to B_CACHE not being set ) and overwriting
3950 * Most callers, including NFS and FFS, need this to
3951 * operate properly either because they assume they
3952 * can issue a read if B_CACHE is not set, or because
3953 * ( for example ) an uncached B_DELWRI might loop due
3954 * to softupdates re-dirtying the buffer. In the latter
3955 * case, B_CACHE is set after the first write completes,
3956 * preventing further loops.
3957 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3958 * above while extending the buffer, we cannot allow the
3959 * buffer to remain with B_CACHE set after the write
3960 * completes or it will represent a corrupt state. To
3961 * deal with this we set B_NOCACHE to scrap the buffer
3964 * We might be able to do something fancy, like setting
3965 * B_CACHE in bwrite() except if B_DELWRI is already set,
3966 * so the below call doesn't set B_CACHE, but that gets real
3967 * confusing. This is much easier.
3970 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3971 bp->b_flags |= B_NOCACHE;
3975 bp->b_flags &= ~B_DONE;
3978 * Buffer is not in-core, create new buffer. The buffer
3979 * returned by getnewbuf() is locked. Note that the returned
3980 * buffer is also considered valid (not marked B_INVAL).
3984 * If the user does not want us to create the buffer, bail out
3987 if (flags & GB_NOCREAT)
3990 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3991 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3992 offset = blkno * bsize;
3993 vmio = vp->v_object != NULL;
3995 maxsize = size + (offset & PAGE_MASK);
3998 /* Do not allow non-VMIO notmapped buffers. */
3999 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4001 maxsize = imax(maxsize, bsize);
4002 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4003 !vn_isdisk(vp, NULL)) {
4004 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4005 KASSERT(error != EOPNOTSUPP,
4006 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4011 return (EJUSTRETURN);
4014 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4016 if (slpflag || slptimeo)
4019 * XXX This is here until the sleep path is diagnosed
4020 * enough to work under very low memory conditions.
4022 * There's an issue on low memory, 4BSD+non-preempt
4023 * systems (eg MIPS routers with 32MB RAM) where buffer
4024 * exhaustion occurs without sleeping for buffer
4025 * reclaimation. This just sticks in a loop and
4026 * constantly attempts to allocate a buffer, which
4027 * hits exhaustion and tries to wakeup bufdaemon.
4028 * This never happens because we never yield.
4030 * The real solution is to identify and fix these cases
4031 * so we aren't effectively busy-waiting in a loop
4032 * until the reclaimation path has cycles to run.
4034 kern_yield(PRI_USER);
4039 * This code is used to make sure that a buffer is not
4040 * created while the getnewbuf routine is blocked.
4041 * This can be a problem whether the vnode is locked or not.
4042 * If the buffer is created out from under us, we have to
4043 * throw away the one we just created.
4045 * Note: this must occur before we associate the buffer
4046 * with the vp especially considering limitations in
4047 * the splay tree implementation when dealing with duplicate
4051 if (gbincore(bo, blkno)) {
4053 bp->b_flags |= B_INVAL;
4054 bufspace_release(bufdomain(bp), maxsize);
4060 * Insert the buffer into the hash, so that it can
4061 * be found by incore.
4063 bp->b_lblkno = blkno;
4064 bp->b_blkno = d_blkno;
4065 bp->b_offset = offset;
4070 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4071 * buffer size starts out as 0, B_CACHE will be set by
4072 * allocbuf() for the VMIO case prior to it testing the
4073 * backing store for validity.
4077 bp->b_flags |= B_VMIO;
4078 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4079 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4080 bp, vp->v_object, bp->b_bufobj->bo_object));
4082 bp->b_flags &= ~B_VMIO;
4083 KASSERT(bp->b_bufobj->bo_object == NULL,
4084 ("ARGH! has b_bufobj->bo_object %p %p\n",
4085 bp, bp->b_bufobj->bo_object));
4086 BUF_CHECK_MAPPED(bp);
4090 bufspace_release(bufdomain(bp), maxsize);
4091 bp->b_flags &= ~B_DONE;
4093 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4095 buf_track(bp, __func__);
4096 KASSERT(bp->b_bufobj == bo,
4097 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4103 * Get an empty, disassociated buffer of given size. The buffer is initially
4107 geteblk(int size, int flags)
4112 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4113 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4114 if ((flags & GB_NOWAIT_BD) &&
4115 (curthread->td_pflags & TDP_BUFNEED) != 0)
4119 bufspace_release(bufdomain(bp), maxsize);
4120 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4125 * Truncate the backing store for a non-vmio buffer.
4128 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4131 if (bp->b_flags & B_MALLOC) {
4133 * malloced buffers are not shrunk
4135 if (newbsize == 0) {
4136 bufmallocadjust(bp, 0);
4137 free(bp->b_data, M_BIOBUF);
4138 bp->b_data = bp->b_kvabase;
4139 bp->b_flags &= ~B_MALLOC;
4143 vm_hold_free_pages(bp, newbsize);
4144 bufspace_adjust(bp, newbsize);
4148 * Extend the backing for a non-VMIO buffer.
4151 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4157 * We only use malloced memory on the first allocation.
4158 * and revert to page-allocated memory when the buffer
4161 * There is a potential smp race here that could lead
4162 * to bufmallocspace slightly passing the max. It
4163 * is probably extremely rare and not worth worrying
4166 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4167 bufmallocspace < maxbufmallocspace) {
4168 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4169 bp->b_flags |= B_MALLOC;
4170 bufmallocadjust(bp, newbsize);
4175 * If the buffer is growing on its other-than-first
4176 * allocation then we revert to the page-allocation
4181 if (bp->b_flags & B_MALLOC) {
4182 origbuf = bp->b_data;
4183 origbufsize = bp->b_bufsize;
4184 bp->b_data = bp->b_kvabase;
4185 bufmallocadjust(bp, 0);
4186 bp->b_flags &= ~B_MALLOC;
4187 newbsize = round_page(newbsize);
4189 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4190 (vm_offset_t) bp->b_data + newbsize);
4191 if (origbuf != NULL) {
4192 bcopy(origbuf, bp->b_data, origbufsize);
4193 free(origbuf, M_BIOBUF);
4195 bufspace_adjust(bp, newbsize);
4199 * This code constitutes the buffer memory from either anonymous system
4200 * memory (in the case of non-VMIO operations) or from an associated
4201 * VM object (in the case of VMIO operations). This code is able to
4202 * resize a buffer up or down.
4204 * Note that this code is tricky, and has many complications to resolve
4205 * deadlock or inconsistent data situations. Tread lightly!!!
4206 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4207 * the caller. Calling this code willy nilly can result in the loss of data.
4209 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4210 * B_CACHE for the non-VMIO case.
4213 allocbuf(struct buf *bp, int size)
4217 if (bp->b_bcount == size)
4220 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4221 panic("allocbuf: buffer too small");
4223 newbsize = roundup2(size, DEV_BSIZE);
4224 if ((bp->b_flags & B_VMIO) == 0) {
4225 if ((bp->b_flags & B_MALLOC) == 0)
4226 newbsize = round_page(newbsize);
4228 * Just get anonymous memory from the kernel. Don't
4229 * mess with B_CACHE.
4231 if (newbsize < bp->b_bufsize)
4232 vfs_nonvmio_truncate(bp, newbsize);
4233 else if (newbsize > bp->b_bufsize)
4234 vfs_nonvmio_extend(bp, newbsize);
4238 desiredpages = (size == 0) ? 0 :
4239 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4241 if (bp->b_flags & B_MALLOC)
4242 panic("allocbuf: VMIO buffer can't be malloced");
4244 * Set B_CACHE initially if buffer is 0 length or will become
4247 if (size == 0 || bp->b_bufsize == 0)
4248 bp->b_flags |= B_CACHE;
4250 if (newbsize < bp->b_bufsize)
4251 vfs_vmio_truncate(bp, desiredpages);
4252 /* XXX This looks as if it should be newbsize > b_bufsize */
4253 else if (size > bp->b_bcount)
4254 vfs_vmio_extend(bp, desiredpages, size);
4255 bufspace_adjust(bp, newbsize);
4257 bp->b_bcount = size; /* requested buffer size. */
4261 extern int inflight_transient_maps;
4263 static struct bio_queue nondump_bios;
4266 biodone(struct bio *bp)
4269 void (*done)(struct bio *);
4270 vm_offset_t start, end;
4272 biotrack(bp, __func__);
4275 * Avoid completing I/O when dumping after a panic since that may
4276 * result in a deadlock in the filesystem or pager code. Note that
4277 * this doesn't affect dumps that were started manually since we aim
4278 * to keep the system usable after it has been resumed.
4280 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4281 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4284 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4285 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4286 bp->bio_flags |= BIO_UNMAPPED;
4287 start = trunc_page((vm_offset_t)bp->bio_data);
4288 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4289 bp->bio_data = unmapped_buf;
4290 pmap_qremove(start, atop(end - start));
4291 vmem_free(transient_arena, start, end - start);
4292 atomic_add_int(&inflight_transient_maps, -1);
4294 done = bp->bio_done;
4296 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4298 bp->bio_flags |= BIO_DONE;
4306 * Wait for a BIO to finish.
4309 biowait(struct bio *bp, const char *wchan)
4313 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4315 while ((bp->bio_flags & BIO_DONE) == 0)
4316 msleep(bp, mtxp, PRIBIO, wchan, 0);
4318 if (bp->bio_error != 0)
4319 return (bp->bio_error);
4320 if (!(bp->bio_flags & BIO_ERROR))
4326 biofinish(struct bio *bp, struct devstat *stat, int error)
4330 bp->bio_error = error;
4331 bp->bio_flags |= BIO_ERROR;
4334 devstat_end_transaction_bio(stat, bp);
4338 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4340 biotrack_buf(struct bio *bp, const char *location)
4343 buf_track(bp->bio_track_bp, location);
4350 * Wait for buffer I/O completion, returning error status. The buffer
4351 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4352 * error and cleared.
4355 bufwait(struct buf *bp)
4357 if (bp->b_iocmd == BIO_READ)
4358 bwait(bp, PRIBIO, "biord");
4360 bwait(bp, PRIBIO, "biowr");
4361 if (bp->b_flags & B_EINTR) {
4362 bp->b_flags &= ~B_EINTR;
4365 if (bp->b_ioflags & BIO_ERROR) {
4366 return (bp->b_error ? bp->b_error : EIO);
4375 * Finish I/O on a buffer, optionally calling a completion function.
4376 * This is usually called from an interrupt so process blocking is
4379 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4380 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4381 * assuming B_INVAL is clear.
4383 * For the VMIO case, we set B_CACHE if the op was a read and no
4384 * read error occurred, or if the op was a write. B_CACHE is never
4385 * set if the buffer is invalid or otherwise uncacheable.
4387 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4388 * initiator to leave B_INVAL set to brelse the buffer out of existence
4389 * in the biodone routine.
4392 bufdone(struct buf *bp)
4394 struct bufobj *dropobj;
4395 void (*biodone)(struct buf *);
4397 buf_track(bp, __func__);
4398 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4401 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4403 runningbufwakeup(bp);
4404 if (bp->b_iocmd == BIO_WRITE)
4405 dropobj = bp->b_bufobj;
4406 /* call optional completion function if requested */
4407 if (bp->b_iodone != NULL) {
4408 biodone = bp->b_iodone;
4409 bp->b_iodone = NULL;
4412 bufobj_wdrop(dropobj);
4415 if (bp->b_flags & B_VMIO) {
4417 * Set B_CACHE if the op was a normal read and no error
4418 * occurred. B_CACHE is set for writes in the b*write()
4421 if (bp->b_iocmd == BIO_READ &&
4422 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4423 !(bp->b_ioflags & BIO_ERROR))
4424 bp->b_flags |= B_CACHE;
4425 vfs_vmio_iodone(bp);
4427 if (!LIST_EMPTY(&bp->b_dep))
4429 if ((bp->b_flags & B_CKHASH) != 0) {
4430 KASSERT(bp->b_iocmd == BIO_READ,
4431 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4432 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4433 (*bp->b_ckhashcalc)(bp);
4436 * For asynchronous completions, release the buffer now. The brelse
4437 * will do a wakeup there if necessary - so no need to do a wakeup
4438 * here in the async case. The sync case always needs to do a wakeup.
4440 if (bp->b_flags & B_ASYNC) {
4441 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4442 (bp->b_ioflags & BIO_ERROR))
4449 bufobj_wdrop(dropobj);
4453 * This routine is called in lieu of iodone in the case of
4454 * incomplete I/O. This keeps the busy status for pages
4458 vfs_unbusy_pages(struct buf *bp)
4465 runningbufwakeup(bp);
4466 if (!(bp->b_flags & B_VMIO))
4469 obj = bp->b_bufobj->bo_object;
4471 for (i = 0; i < bp->b_npages; i++) {
4473 if (m == bogus_page) {
4474 if (bogus == false) {
4476 VM_OBJECT_RLOCK(obj);
4478 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4480 panic("vfs_unbusy_pages: page missing\n");
4482 if (buf_mapped(bp)) {
4483 BUF_CHECK_MAPPED(bp);
4484 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4485 bp->b_pages, bp->b_npages);
4487 BUF_CHECK_UNMAPPED(bp);
4492 VM_OBJECT_RUNLOCK(obj);
4493 vm_object_pip_wakeupn(obj, bp->b_npages);
4497 * vfs_page_set_valid:
4499 * Set the valid bits in a page based on the supplied offset. The
4500 * range is restricted to the buffer's size.
4502 * This routine is typically called after a read completes.
4505 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4510 * Compute the end offset, eoff, such that [off, eoff) does not span a
4511 * page boundary and eoff is not greater than the end of the buffer.
4512 * The end of the buffer, in this case, is our file EOF, not the
4513 * allocation size of the buffer.
4515 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4516 if (eoff > bp->b_offset + bp->b_bcount)
4517 eoff = bp->b_offset + bp->b_bcount;
4520 * Set valid range. This is typically the entire buffer and thus the
4524 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4528 * vfs_page_set_validclean:
4530 * Set the valid bits and clear the dirty bits in a page based on the
4531 * supplied offset. The range is restricted to the buffer's size.
4534 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4536 vm_ooffset_t soff, eoff;
4539 * Start and end offsets in buffer. eoff - soff may not cross a
4540 * page boundary or cross the end of the buffer. The end of the
4541 * buffer, in this case, is our file EOF, not the allocation size
4545 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4546 if (eoff > bp->b_offset + bp->b_bcount)
4547 eoff = bp->b_offset + bp->b_bcount;
4550 * Set valid range. This is typically the entire buffer and thus the
4554 vm_page_set_validclean(
4556 (vm_offset_t) (soff & PAGE_MASK),
4557 (vm_offset_t) (eoff - soff)
4563 * Acquire a shared busy on all pages in the buf.
4566 vfs_busy_pages_acquire(struct buf *bp)
4570 for (i = 0; i < bp->b_npages; i++)
4571 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4575 vfs_busy_pages_release(struct buf *bp)
4579 for (i = 0; i < bp->b_npages; i++)
4580 vm_page_sunbusy(bp->b_pages[i]);
4584 * This routine is called before a device strategy routine.
4585 * It is used to tell the VM system that paging I/O is in
4586 * progress, and treat the pages associated with the buffer
4587 * almost as being exclusive busy. Also the object paging_in_progress
4588 * flag is handled to make sure that the object doesn't become
4591 * Since I/O has not been initiated yet, certain buffer flags
4592 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4593 * and should be ignored.
4596 vfs_busy_pages(struct buf *bp, int clear_modify)
4604 if (!(bp->b_flags & B_VMIO))
4607 obj = bp->b_bufobj->bo_object;
4608 foff = bp->b_offset;
4609 KASSERT(bp->b_offset != NOOFFSET,
4610 ("vfs_busy_pages: no buffer offset"));
4611 if ((bp->b_flags & B_CLUSTER) == 0) {
4612 vm_object_pip_add(obj, bp->b_npages);
4613 vfs_busy_pages_acquire(bp);
4615 if (bp->b_bufsize != 0)
4616 vfs_setdirty_range(bp);
4618 for (i = 0; i < bp->b_npages; i++) {
4620 vm_page_assert_sbusied(m);
4623 * When readying a buffer for a read ( i.e
4624 * clear_modify == 0 ), it is important to do
4625 * bogus_page replacement for valid pages in
4626 * partially instantiated buffers. Partially
4627 * instantiated buffers can, in turn, occur when
4628 * reconstituting a buffer from its VM backing store
4629 * base. We only have to do this if B_CACHE is
4630 * clear ( which causes the I/O to occur in the
4631 * first place ). The replacement prevents the read
4632 * I/O from overwriting potentially dirty VM-backed
4633 * pages. XXX bogus page replacement is, uh, bogus.
4634 * It may not work properly with small-block devices.
4635 * We need to find a better way.
4638 pmap_remove_write(m);
4639 vfs_page_set_validclean(bp, foff, m);
4640 } else if (vm_page_all_valid(m) &&
4641 (bp->b_flags & B_CACHE) == 0) {
4642 bp->b_pages[i] = bogus_page;
4645 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4647 if (bogus && buf_mapped(bp)) {
4648 BUF_CHECK_MAPPED(bp);
4649 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4650 bp->b_pages, bp->b_npages);
4655 * vfs_bio_set_valid:
4657 * Set the range within the buffer to valid. The range is
4658 * relative to the beginning of the buffer, b_offset. Note that
4659 * b_offset itself may be offset from the beginning of the first
4663 vfs_bio_set_valid(struct buf *bp, int base, int size)
4668 if (!(bp->b_flags & B_VMIO))
4672 * Fixup base to be relative to beginning of first page.
4673 * Set initial n to be the maximum number of bytes in the
4674 * first page that can be validated.
4676 base += (bp->b_offset & PAGE_MASK);
4677 n = PAGE_SIZE - (base & PAGE_MASK);
4680 * Busy may not be strictly necessary here because the pages are
4681 * unlikely to be fully valid and the vnode lock will synchronize
4682 * their access via getpages. It is grabbed for consistency with
4683 * other page validation.
4685 vfs_busy_pages_acquire(bp);
4686 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4690 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4695 vfs_busy_pages_release(bp);
4701 * If the specified buffer is a non-VMIO buffer, clear the entire
4702 * buffer. If the specified buffer is a VMIO buffer, clear and
4703 * validate only the previously invalid portions of the buffer.
4704 * This routine essentially fakes an I/O, so we need to clear
4705 * BIO_ERROR and B_INVAL.
4707 * Note that while we only theoretically need to clear through b_bcount,
4708 * we go ahead and clear through b_bufsize.
4711 vfs_bio_clrbuf(struct buf *bp)
4713 int i, j, mask, sa, ea, slide;
4715 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4719 bp->b_flags &= ~B_INVAL;
4720 bp->b_ioflags &= ~BIO_ERROR;
4721 vfs_busy_pages_acquire(bp);
4722 sa = bp->b_offset & PAGE_MASK;
4724 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4725 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4726 ea = slide & PAGE_MASK;
4729 if (bp->b_pages[i] == bogus_page)
4732 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4733 if ((bp->b_pages[i]->valid & mask) == mask)
4735 if ((bp->b_pages[i]->valid & mask) == 0)
4736 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4738 for (; sa < ea; sa += DEV_BSIZE, j++) {
4739 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4740 pmap_zero_page_area(bp->b_pages[i],
4745 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4746 roundup2(ea - sa, DEV_BSIZE));
4748 vfs_busy_pages_release(bp);
4753 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4758 if (buf_mapped(bp)) {
4759 BUF_CHECK_MAPPED(bp);
4760 bzero(bp->b_data + base, size);
4762 BUF_CHECK_UNMAPPED(bp);
4763 n = PAGE_SIZE - (base & PAGE_MASK);
4764 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4768 pmap_zero_page_area(m, base & PAGE_MASK, n);
4777 * Update buffer flags based on I/O request parameters, optionally releasing the
4778 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4779 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4780 * I/O). Otherwise the buffer is released to the cache.
4783 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4786 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4787 ("buf %p non-VMIO noreuse", bp));
4789 if ((ioflag & IO_DIRECT) != 0)
4790 bp->b_flags |= B_DIRECT;
4791 if ((ioflag & IO_EXT) != 0)
4792 bp->b_xflags |= BX_ALTDATA;
4793 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4794 bp->b_flags |= B_RELBUF;
4795 if ((ioflag & IO_NOREUSE) != 0)
4796 bp->b_flags |= B_NOREUSE;
4804 vfs_bio_brelse(struct buf *bp, int ioflag)
4807 b_io_dismiss(bp, ioflag, true);
4811 vfs_bio_set_flags(struct buf *bp, int ioflag)
4814 b_io_dismiss(bp, ioflag, false);
4818 * vm_hold_load_pages and vm_hold_free_pages get pages into
4819 * a buffers address space. The pages are anonymous and are
4820 * not associated with a file object.
4823 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4829 BUF_CHECK_MAPPED(bp);
4831 to = round_page(to);
4832 from = round_page(from);
4833 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4835 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4837 * note: must allocate system pages since blocking here
4838 * could interfere with paging I/O, no matter which
4841 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4842 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4844 pmap_qenter(pg, &p, 1);
4845 bp->b_pages[index] = p;
4847 bp->b_npages = index;
4850 /* Return pages associated with this buf to the vm system */
4852 vm_hold_free_pages(struct buf *bp, int newbsize)
4856 int index, newnpages;
4858 BUF_CHECK_MAPPED(bp);
4860 from = round_page((vm_offset_t)bp->b_data + newbsize);
4861 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4862 if (bp->b_npages > newnpages)
4863 pmap_qremove(from, bp->b_npages - newnpages);
4864 for (index = newnpages; index < bp->b_npages; index++) {
4865 p = bp->b_pages[index];
4866 bp->b_pages[index] = NULL;
4867 vm_page_unwire_noq(p);
4870 bp->b_npages = newnpages;
4874 * Map an IO request into kernel virtual address space.
4876 * All requests are (re)mapped into kernel VA space.
4877 * Notice that we use b_bufsize for the size of the buffer
4878 * to be mapped. b_bcount might be modified by the driver.
4880 * Note that even if the caller determines that the address space should
4881 * be valid, a race or a smaller-file mapped into a larger space may
4882 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4883 * check the return value.
4885 * This function only works with pager buffers.
4888 vmapbuf(struct buf *bp, int mapbuf)
4893 if (bp->b_bufsize < 0)
4895 prot = VM_PROT_READ;
4896 if (bp->b_iocmd == BIO_READ)
4897 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4898 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4899 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4900 btoc(MAXPHYS))) < 0)
4902 bp->b_npages = pidx;
4903 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4904 if (mapbuf || !unmapped_buf_allowed) {
4905 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4906 bp->b_data = bp->b_kvabase + bp->b_offset;
4908 bp->b_data = unmapped_buf;
4913 * Free the io map PTEs associated with this IO operation.
4914 * We also invalidate the TLB entries and restore the original b_addr.
4916 * This function only works with pager buffers.
4919 vunmapbuf(struct buf *bp)
4923 npages = bp->b_npages;
4925 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4926 vm_page_unhold_pages(bp->b_pages, npages);
4928 bp->b_data = unmapped_buf;
4932 bdone(struct buf *bp)
4936 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4938 bp->b_flags |= B_DONE;
4944 bwait(struct buf *bp, u_char pri, const char *wchan)
4948 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4950 while ((bp->b_flags & B_DONE) == 0)
4951 msleep(bp, mtxp, pri, wchan, 0);
4956 bufsync(struct bufobj *bo, int waitfor)
4959 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4963 bufstrategy(struct bufobj *bo, struct buf *bp)
4969 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4970 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4971 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4972 i = VOP_STRATEGY(vp, bp);
4973 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4977 * Initialize a struct bufobj before use. Memory is assumed zero filled.
4980 bufobj_init(struct bufobj *bo, void *private)
4982 static volatile int bufobj_cleanq;
4985 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
4986 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
4987 bo->bo_private = private;
4988 TAILQ_INIT(&bo->bo_clean.bv_hd);
4989 TAILQ_INIT(&bo->bo_dirty.bv_hd);
4993 bufobj_wrefl(struct bufobj *bo)
4996 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4997 ASSERT_BO_WLOCKED(bo);
5002 bufobj_wref(struct bufobj *bo)
5005 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5012 bufobj_wdrop(struct bufobj *bo)
5015 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5017 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5018 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5019 bo->bo_flag &= ~BO_WWAIT;
5020 wakeup(&bo->bo_numoutput);
5026 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5030 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5031 ASSERT_BO_WLOCKED(bo);
5033 while (bo->bo_numoutput) {
5034 bo->bo_flag |= BO_WWAIT;
5035 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5036 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5044 * Set bio_data or bio_ma for struct bio from the struct buf.
5047 bdata2bio(struct buf *bp, struct bio *bip)
5050 if (!buf_mapped(bp)) {
5051 KASSERT(unmapped_buf_allowed, ("unmapped"));
5052 bip->bio_ma = bp->b_pages;
5053 bip->bio_ma_n = bp->b_npages;
5054 bip->bio_data = unmapped_buf;
5055 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5056 bip->bio_flags |= BIO_UNMAPPED;
5057 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5058 PAGE_SIZE == bp->b_npages,
5059 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5060 (long long)bip->bio_length, bip->bio_ma_n));
5062 bip->bio_data = bp->b_data;
5068 * The MIPS pmap code currently doesn't handle aliased pages.
5069 * The VIPT caches may not handle page aliasing themselves, leading
5070 * to data corruption.
5072 * As such, this code makes a system extremely unhappy if said
5073 * system doesn't support unaliasing the above situation in hardware.
5074 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5075 * this feature at build time, so it has to be handled in software.
5077 * Once the MIPS pmap/cache code grows to support this function on
5078 * earlier chips, it should be flipped back off.
5081 static int buf_pager_relbuf = 1;
5083 static int buf_pager_relbuf = 0;
5085 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5086 &buf_pager_relbuf, 0,
5087 "Make buffer pager release buffers after reading");
5090 * The buffer pager. It uses buffer reads to validate pages.
5092 * In contrast to the generic local pager from vm/vnode_pager.c, this
5093 * pager correctly and easily handles volumes where the underlying
5094 * device block size is greater than the machine page size. The
5095 * buffer cache transparently extends the requested page run to be
5096 * aligned at the block boundary, and does the necessary bogus page
5097 * replacements in the addends to avoid obliterating already valid
5100 * The only non-trivial issue is that the exclusive busy state for
5101 * pages, which is assumed by the vm_pager_getpages() interface, is
5102 * incompatible with the VMIO buffer cache's desire to share-busy the
5103 * pages. This function performs a trivial downgrade of the pages'
5104 * state before reading buffers, and a less trivial upgrade from the
5105 * shared-busy to excl-busy state after the read.
5108 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5109 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5110 vbg_get_blksize_t get_blksize)
5117 vm_ooffset_t la, lb, poff, poffe;
5119 int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5122 object = vp->v_object;
5125 la = IDX_TO_OFF(ma[count - 1]->pindex);
5126 if (la >= object->un_pager.vnp.vnp_size)
5127 return (VM_PAGER_BAD);
5130 * Change the meaning of la from where the last requested page starts
5131 * to where it ends, because that's the end of the requested region
5132 * and the start of the potential read-ahead region.
5135 lpart = la > object->un_pager.vnp.vnp_size;
5136 bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5139 * Calculate read-ahead, behind and total pages.
5142 lb = IDX_TO_OFF(ma[0]->pindex);
5143 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5145 if (rbehind != NULL)
5147 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5148 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5149 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5154 VM_CNT_INC(v_vnodein);
5155 VM_CNT_ADD(v_vnodepgsin, pgsin);
5157 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5158 != 0) ? GB_UNMAPPED : 0;
5160 for (i = 0; i < count; i++)
5161 vm_page_busy_downgrade(ma[i]);
5164 for (i = 0; i < count; i++) {
5168 * Pages are shared busy and the object lock is not
5169 * owned, which together allow for the pages'
5170 * invalidation. The racy test for validity avoids
5171 * useless creation of the buffer for the most typical
5172 * case when invalidation is not used in redo or for
5173 * parallel read. The shared->excl upgrade loop at
5174 * the end of the function catches the race in a
5175 * reliable way (protected by the object lock).
5177 if (vm_page_all_valid(m))
5180 poff = IDX_TO_OFF(m->pindex);
5181 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5182 for (; poff < poffe; poff += bsize) {
5183 lbn = get_lblkno(vp, poff);
5188 bsize = get_blksize(vp, lbn);
5189 error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5193 if (LIST_EMPTY(&bp->b_dep)) {
5195 * Invalidation clears m->valid, but
5196 * may leave B_CACHE flag if the
5197 * buffer existed at the invalidation
5198 * time. In this case, recycle the
5199 * buffer to do real read on next
5200 * bread() after redo.
5202 * Otherwise B_RELBUF is not strictly
5203 * necessary, enable to reduce buf
5206 if (buf_pager_relbuf ||
5207 !vm_page_all_valid(m))
5208 bp->b_flags |= B_RELBUF;
5210 bp->b_flags &= ~B_NOCACHE;
5216 KASSERT(1 /* racy, enable for debugging */ ||
5217 vm_page_all_valid(m) || i == count - 1,
5218 ("buf %d %p invalid", i, m));
5219 if (i == count - 1 && lpart) {
5220 if (!vm_page_none_valid(m) &&
5221 !vm_page_all_valid(m))
5222 vm_page_zero_invalid(m, TRUE);
5228 VM_OBJECT_WLOCK(object);
5230 for (i = 0; i < count; i++) {
5231 vm_page_sunbusy(ma[i]);
5232 ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5235 * Since the pages were only sbusy while neither the
5236 * buffer nor the object lock was held by us, or
5237 * reallocated while vm_page_grab() slept for busy
5238 * relinguish, they could have been invalidated.
5239 * Recheck the valid bits and re-read as needed.
5241 * Note that the last page is made fully valid in the
5242 * read loop, and partial validity for the page at
5243 * index count - 1 could mean that the page was
5244 * invalidated or removed, so we must restart for
5247 if (!vm_page_all_valid(ma[i]))
5250 VM_OBJECT_WUNLOCK(object);
5251 if (redo && error == 0)
5253 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5256 #include "opt_ddb.h"
5258 #include <ddb/ddb.h>
5260 /* DDB command to show buffer data */
5261 DB_SHOW_COMMAND(buffer, db_show_buffer)
5264 struct buf *bp = (struct buf *)addr;
5265 #ifdef FULL_BUF_TRACKING
5270 db_printf("usage: show buffer <addr>\n");
5274 db_printf("buf at %p\n", bp);
5275 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5276 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5277 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5278 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5279 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5280 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5282 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5283 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5284 "b_vp = %p, b_dep = %p\n",
5285 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5286 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5287 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5288 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5289 bp->b_kvabase, bp->b_kvasize);
5292 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5293 for (i = 0; i < bp->b_npages; i++) {
5297 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5299 (u_long)VM_PAGE_TO_PHYS(m));
5301 db_printf("( ??? )");
5302 if ((i + 1) < bp->b_npages)
5307 BUF_LOCKPRINTINFO(bp);
5308 #if defined(FULL_BUF_TRACKING)
5309 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5311 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5312 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5313 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5315 db_printf(" %2u: %s\n", j,
5316 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5318 #elif defined(BUF_TRACKING)
5319 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5324 DB_SHOW_COMMAND(bufqueues, bufqueues)
5326 struct bufdomain *bd;
5331 db_printf("bqempty: %d\n", bqempty.bq_len);
5333 for (i = 0; i < buf_domains; i++) {
5335 db_printf("Buf domain %d\n", i);
5336 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5337 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5338 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5340 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5341 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5342 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5343 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5344 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5346 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5347 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5348 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5349 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5352 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5353 total += bp->b_bufsize;
5354 db_printf("\tcleanq count\t%d (%ld)\n",
5355 bd->bd_cleanq->bq_len, total);
5357 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5358 total += bp->b_bufsize;
5359 db_printf("\tdirtyq count\t%d (%ld)\n",
5360 bd->bd_dirtyq.bq_len, total);
5361 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5362 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5363 db_printf("\tCPU ");
5364 for (j = 0; j <= mp_maxid; j++)
5365 db_printf("%d, ", bd->bd_subq[j].bq_len);
5369 for (j = 0; j < nbuf; j++)
5370 if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5372 total += buf[j].b_bufsize;
5374 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5377 for (j = 0; j < nbuf; j++)
5378 if (buf[j].b_domain == i) {
5380 total += buf[j].b_bufsize;
5382 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5386 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5391 for (i = 0; i < nbuf; i++) {
5393 if (BUF_ISLOCKED(bp)) {
5394 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5402 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5408 db_printf("usage: show vnodebufs <addr>\n");
5411 vp = (struct vnode *)addr;
5412 db_printf("Clean buffers:\n");
5413 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5414 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5417 db_printf("Dirty buffers:\n");
5418 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5419 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5424 DB_COMMAND(countfreebufs, db_coundfreebufs)
5427 int i, used = 0, nfree = 0;
5430 db_printf("usage: countfreebufs\n");
5434 for (i = 0; i < nbuf; i++) {
5436 if (bp->b_qindex == QUEUE_EMPTY)
5442 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5444 db_printf("numfreebuffers is %d\n", numfreebuffers);